Let's talk

LinkedIn

Twitter

Medium

Youtube

Download Guide

<h1>Download 
the guide</h1>
A three-step guide about how to make personalisation a reality.

Datatonic and Intel Whitepaper: How to be Hyper-Relevant

Generative AI + LLMOps: Applications in Production

Generative AI 101: GenAI on Google Cloud

Responsible AI: The Role of Data and Model Cards

At our <a href="https://datatonic.com/events/driving-analytics-adoption-with-admiral-insurance/" target="_blank" rel="noopener">Driving Analytics Adoption</a> webinar, Datatonic had the pleasure of welcoming data experts, <a href="https://www.linkedin.com/in/rachel-purchase-4b738084/" target="_blank" rel="noopener">Rachel Purchase</a>, Head of Data Product, Insights and Data Science at Admiral and <a href="https://www.linkedin.com/in/hldavies1/" target="_blank" rel="noopener">Hannah Davies</a>, Head of Data Academy at Admiral.
Along with Datatonic’s Head of Analytics, <a href="https://www.linkedin.com/in/rahul-prasad-5265592a/" target="_blank" rel="noopener">Rahul Prasad</a> and Datatonic Academy Co-Founder, <a href="https://www.linkedin.com/in/ambercella/" target="_blank" rel="noopener">Amber Cella</a>, they discussed the various ways to increase analytics adoption at scale across your business.
Analytics Adoption
Most companies know that staying up to date with the latest tools is important. Working with the best tools available makes work easier and more efficient for tech teams. Companies need to make the most of the useful data available to them, and one of the first steps towards this is better tooling that can accelerate the time to insights. 
However, a recent Gartner, Inc. <a href="https://www.gartner.com/smarterwithgartner/take-your-analytics-maturity-to-the-next-level">survey</a> found that 87.5% of organisations classified their data maturity as low, with many large organisations still reliant on spreadsheets for analysis. Investing in new tooling is only worth it if business users are enabled to use the tools available to them. While 63% of businesses are increasing their digital transformation budgets, the adoption rate for BI and analytics tools is still only around 25-35%.
During our webinar, the audience gave us great insights into some of the common questions businesses might have when looking to increase analytics adoption, ranging from the best tools for documentation to getting company and HR backing for new organisational structures. Read to find out what our experts said.
<h3>Q + A: Your Questions Answered</h3>
&nbsp;
<ol>
<li> What has been the biggest challenge in the Admiral analytics adoption journey so far?</li>
</ol>
&nbsp;
Rachel: For us, it&#8217;s been about hearts and minds across the whole journey. If we go right back to the beginning, for example, when we were starting, we were pretty much a completely on-prem-based company so we had all of our own data centres. We were really excited but we had to go to security and say we&#8217;re going to the cloud and we&#8217;re going to put all of our data there.
For businesses that are yet to have that conversation, it&#8217;s all about people: convincing security and talking to your end users who are attached to their existing tool stack. You&#8217;ve got to think about people if you want to drive adoption. I think that&#8217;s our biggest challenge and we probably learned too late. It took us two and a bit years to start the academy and that was probably two in a bit years too late.
Hannah: It&#8217;s about recognising that you can get excited about new tooling but not so excited that you lose your objectives. Making sure that you&#8217;re clear on what you want at the end and thinking about that journey to getting people through it is key. It&#8217;s great to win people over with a fancy tool, but when they&#8217;re attached to the tooling that they have today or the way that they work today, it can be a hard journey. Your teams have to be at the heart of every decision you make and come into it, thinking what would I do in that situation, what would I want from it, and how do we leverage that to make it successful?
&nbsp;
<ol start="2">
<li> Should companies have a centralised BI development team as part of their CoE? If so, how do you manage this team and their long-term work versus the ad-hoc and quick turnaround data analysis?</li>
</ol>
&nbsp;
Rahul: I&#8217;ve definitely seen some clients wanting to have a centralised BI team sitting under a Center of Excellence and especially an Analytics Center of Excellence. If you have a mature BI team who has a very well-defined set of processes, it&#8217;s not too difficult. If you&#8217;re just starting out, you need a very well-structured governance plan around how you would manage this team. 
What I&#8217;ve seen work quite nicely is having someone from the analytics team actually sitting as part of the Center of Excellence so the communication channel between those two groups or teams is almost seamless.
For ad hoc and long-term reporting, this depends on the project. If it&#8217;s a pure BI project, I don&#8217;t see why it would be a problem if you&#8217;ve given the team autonomy around how they can go and deliver these solutions. Ad hoc analytics can be a slightly tricky one, especially on the BI side of things as we now have self-serve and also some of the large organisations also want to retain more ad hoc analysis. It’s a tough decision to say “You can only do one or the other”. 
You have to find a balance between how you would do ad hoc analysis and at the same time give people regular reporting. The concept of having golden reports has worked quite nicely for me in the past, having well-vetted reporting solutions that have gone through proper code review, approval from various team members, etc. This is like your golden area where you know the numbers are not going to change and then you can use this for ad hoc analysis.
&nbsp;
Rachel: At Admiral, we have a centralised BI team as part of the CoE. The first important thing to consider if you&#8217;re going to do that is “What&#8217;s their purpose? Why are they there instead of sitting out in your business?”. For us, there are a few reasons but one of the big reasons is that we&#8217;re a financial services company. This means that we&#8217;re very heavily regulated and we have to do a lot of regulatory reporting. There&#8217;s a significant amount of rigour that goes around that and it spans the entire business, not just a single business unit.
We assign that reporting to that central team, which means it is alongside best practices and those longer-term projects that the central team does too. One of the risks of course if you have a central team just focused on best practices is that they get out of date with their own skills because they&#8217;re no longer doing complex BI work. 
By its nature, the regulatory reporting and critical work that we&#8217;re assigning to them can be very ad hoc with last-minute requests. That can be disruptive and so you just have to think about your ways of working carefully to make sure that you factor in an expected level of disruption to these sorts of long-term projects.
&nbsp;
<ol start="3">
<li> Do you need to bring in external trainers to set up a Data Academy?</li>
</ol>
&nbsp;
Amber: From our side, a big part of what we do at Datatonic Academy is enabling long-term sustainable upskilling solutions. There are lots of ways of approaching that; one way we do it is through “train the trainer” programs. We enable companies to have their own experts internally. Those people might not be full-time trainers; there are definitely other ways of doing it but it&#8217;s adding a string to the bow of your employees who have either a natural interest or a natural talent in teaching, training, and knowledge sharing.
The investment certainly is more front-loaded if you&#8217;re enabling those experts and those instructors but you don&#8217;t need to hire a dedicated team of instructors. Some businesses want an in-between solution where companies like ourselves come in and work with you over twelve months, two years, or three years to offer regular refresher training to blend tailored training courses that you know use all of your tools and use cases. 
&nbsp;
<ol start="4">
<li> How can companies sell a data academy to their HR team? The investment can be quite big and far-reaching. What&#8217;s the pitch to get the wider team members on your side? </li>
</ol>
&nbsp;
Hannah: For us, It didn&#8217;t take a lot of investment upfront because we found great people, already at Admiral, willing to help out; we do have trainers inside of the team but we also utilise people across the business with secondments into the Data Academy part-time. We have a business where new starters start all of the time and people are already doing some of that training for us, so that space of people who could do it was quite large. 
There&#8217;s so much out there in terms of content such as great webinars like this but also really technical content on YouTube. There&#8217;s lots of material out there but there are also probably some great talented people across the business who might just need a bit of support in being confident to deliver something; that&#8217;s not always a dedicated trainer. It could just be somebody who loves public speaking or being able to coach somebody through a problem. 
Learning and Development have always been a key part of our business. We have a centralised Learning and Development team that covers the whole business as well so it&#8217;s been well-received here because we&#8217;re people-orientated and focused on learning and developing careers. So, the Data Academy wasn&#8217;t a hard sell, coupled with the fact that it&#8217;s not a huge team of 20 people that you&#8217;re trying to bring in. 
If you can focus on achieving value fast, and showing the benefits outweigh the cost, you just need to find a sponsor in the business who supports it and thinks about the benefits in terms of value who can help get you over the line.
&nbsp;
<ol start="5">
<li> What are the tools for documentation you would recommend to create a good Center of Excellence?</li>
</ol>
&nbsp;
Rahul: We have found Lucidchart working very nicely. We also use Miro boards. There is also something called Lucidspark as well that&#8217;s more for wireframing and building storyboards. Most organisations also have a front-end or user experience team as well so they might be using things like design wireframing tools. Those are some of my personal favourites but I would always recommend reaching out to some of your team members to see what they are already using.
Rachel: We’re using very different tools for our primary documentation; we use Confluence because it links into Jira and so it works through our flows of work and then we use Collibra which works very well.
&nbsp;
<ol start="6">
<li> What is the tool stack you use in the Datatonic Academy or do you adapt to what the client company has?</li>
</ol>
&nbsp;
Amber: Our Academy trainers have experience with Google Cloud, AWS, Azure and related technologies (including Cisco, Juniper, dbt, etc.). We offer a number of standard courses across these and provide customisation to maximise relevance to your business. This includes building in business-specific use cases and hands-on projects in addition to emphasising or de-emphasising topics.
We also offer fully custom courses and learning paths which we build to fit your requirements. Once we have scoped your training needs and understood your tech stack, we adapt our offering to fit. Where you require training that’s outside of our specialisms (data + AI), we have a number of partners we work with who can provide high-quality instructor-led training (e.g. for Project Management or related non-technical learning paths).
&nbsp;
Summary
To be successful in driving adoption, you need to be holistic in your approach by understanding it from the perspective of people, processes, and technology. 
New tooling can be a game-changer for improving your analytics capabilities, but it will only have a significant impact if you are able to drive adoption at scale across your teams. This can be done through:
&nbsp;
<ul>
<li style="font-weight: 400;" aria-level="1">Adopting an Analytics Center of Excellence to implement best practices and increase governance</li>
<li style="font-weight: 400;" aria-level="1">Using a training partner such as Datatonic Academy to upskill and increase familiarity with new tools</li>
<li style="font-weight: 400;" aria-level="1">Developing an internal Data Academy to ensure constant upskilling support </li>
</ul>
&nbsp;
Looking for support on driving adoption at scale? Watch our webinar <a href="https://datatonic.com/events/driving-analytics-adoption-with-admiral-insurance/" target="_blank" rel="noopener">here</a> if you haven’t already!
Find out more about training and upskilling at <a href="https://datatonic.academy/" target="_blank" rel="noopener">datatonic.academy</a>, and get in touch <a href="https://datatonic.com/contact/" target="_blank" rel="noopener">here</a> if you are considering developing an Analytics Center of Excellence, or for any other analytics projects.

Driving Analytics Adoption Q+A: Your Questions Answered

<h3>Introduction</h3>
In recent years, many companies have looked to AI and Machine Learning as key technologies to drive growth by creating new revenue streams, reducing operating costs, or mitigating risks. However, leveraging these technologies can be difficult and costly without the right tools, best practices, or expertise. 
Fortunately, many tools have emerged to address these challenges in recent years, such as Google Cloud&#8217;s <a href="https://cloud.google.com/vertex-ai">Vertex AI</a>. These powerful ecosystems can enable companies to harness the full potential of Data Science (DS) and Machine Learning (ML). They can also allow companies to streamline their Machine Learning operations (MLOps), and create incremental value from ML. But, with so many options available, it can be overwhelming for companies to determine which tool is the best fit for their specific needs.
As <a href="https://datatonic.com/insights/datatonic-google-cloud-specialization-partner-award-machine-learning/">Google Cloud’s Machine Learning Partner of the Year</a>, Datatonic has deep expertise in Vertex AI and Google Cloud. In this blog, we will explore how Datatonic leverages its expertise to advise clients on how to use Vertex AI to onboard strategic ML use cases over a two-week-long engagement. 
First, we will discuss how our Vertex AI Assessment package introduces Google Cloud + Vertex AI. Then, we will look at how we support companies in getting started with Vertex AI and allow them to assess its suitability for their ML use cases. Finally, we’ll look at a client case study and will examine how our expertise enabled an informed decision between AWS Sagemaker and Vertex AI for its ML platform.
<h3>Vertex AI Assessment &#8211; Enablement + Hands-on Workshops</h3>
​​Datatonic’s Vertex AI Assessment is designed to provide businesses with a detailed view of Vertex AI and to help them to assess whether or not it is the right fit for their organisation. The engagement primarily covers how Data Scientists and Engineers can use Vertex AI tools to leverage their data and produce state-of-the-art Machine Learning systems. In addition to this, clients are also shown how ML systems can be deployed, tracked and managed using <a href="https://cloud.google.com/vertex-ai/docs/pipelines/introduction">Vertex AI Pipelines</a>.
Workshops run over two weeks and include the initial set-up of Vertex AI infrastructure, a series of deep dive discussions on a curated list of Vertex AI services, and hands-on time for the client to try out Vertex AI. Data Scientists and Machine Learning Engineers from Datatonic are also available throughout the two weeks to support clients as they explore the tools.
Upon completion of the workshops, Datatonic provides expert advice regarding the suitability of Vertex for the client’s needs and assists them in comparing it to any alternative tools they may be considering.
<h3>Case Study</h3>
<h4>Client Background</h4>
The business provides a B2B sales platform that helps companies find their next sales opportunities by matching clients with more relevant customers, allowing for more business to be completed with fewer interactions. The platform accesses high-dimensional databases and harnesses the power of AI and Machine Learning methods to identify sales leads efficiently and reliably. 
<h4>Client Challenge</h4>
Our client has a talented and advanced Data Science research team that is creating innovative solutions to give them an edge over their competitors. With a hybrid cross-cloud set-up, it lacked a standardised process of taking a Data Science solution to production using Cloud technologies. As a result, despite developing Artificial Intelligence and Machine Learning models to improve and automate decision-making, there was still a large manual overhead. These are common challenges among many businesses that are looking to adopt or have already adopted Machine Learning. 
A direct consequence of these challenges is a slow time to production and low ROI from ML initiatives. Similarly, the difficulties involved in monitoring and maintaining these models in production means that the quality of these models may degrade over time. For these reasons, the client was keen to build a platform that can cater for its Data Science and MLOps needs, allowing for a more efficient, reliable and standardised method of productionising cutting-edge Machine Learning solutions, keeping them ahead of the competition. 
While Vertex AI on Google Cloud was identified as a platform where the value of Data Science work could be realised, the client also considered Amazon Web Services’ Sagemaker as an alternative. Datatonic provided a comprehensive evaluation package of the Vertex AI platform, to allow the client to make an informed decision on how to best continue its AI practice.
<h3>Our Solution</h3>
Google Cloud and Vertex AI provide tooling for the entire lifecycle of a Machine Learning solution. This ranges from data ingestion, exploration and model development, all the way to Machine Learning pipelines and model deployment. As the client already has expertise in Data Science, emphasis was placed on how a Data Scientist would use Google Cloud on a day-to-day basis and leverage the relevant services to improve and automate their workflow. With an emphasis on improving existing workflows, deep dives into Vertex Pipelines and MLOps were also at the forefront of the engagement. 
A series of workshops were delivered on the relevant offerings of Google Cloud, including deep dives into the services, tutorials and hands-on lab activities. The first group of workshops discussed “An Introduction to Data Science on Google Cloud”, emphasising how Data Scientists could leverage Google Cloud solutions to reduce the manual overhead in their existing workflows. Following Google Cloud’s recommended Data Science workflow, pre-packaged solutions were explored first, followed by no-code and low-code solutions, and then more advanced custom training options. In addition to modelling capabilities, additional Vertex AI services included: Managed Notebooks, Managed Datasets and Matching Engine, and more.
<h3>Getting Started</h3>
Initially, the client explored the various methods for ingesting and storing data in Google Cloud and the various tools and databases available depending on the data type and use case requirements. This involved covering best practices for <a href="https://cloud.google.com/learn/what-is-etl">ETL</a> on Google Cloud and how to automate data processing so that Data Scientists can spend less time on this aspect and more time generating insights and delivering value from Machine Learning. 
Vertex AI Workbench
Now that the data was transformed and in place, the client’s Data Scientists were eager to get hands-on with the data. Before the engagement, much of the Data Scientist’s work was undertaken locally and thus we were keen to show the power and flexibility of <a href="https://cloud.google.com/vertex-ai-workbench">Vertex AI Workbench</a>. 
Vertex AI Workbench is ideal for end-to-end Data Science workflows; you can implement code for data preparation, exploratory data analysis, custom model training and much more. This service is a Jupyter notebook-based development environment which allows for seamless interaction and integration with Vertex AI and other Google Cloud services, giving full control to your Data Scientists. Within Vertex AI Workbench, there are two options for development: Managed Notebooks and User-Managed Notebooks. 
Managed Notebooks are Google-managed environments that allow for seamless integration with <a href="https://cloud.google.com/bigquery">BigQuery</a> and <a href="https://cloud.google.com/storage">Google Cloud Storage</a>, meaning that Data Scientists do not need to leave the JupyterLab environment to view and explore the data. Managed Notebooks offer further flexibility including the management of hardware and framework for your instance. This is incredibly powerful, as at your fingertips within your notebook interface you can scale how many CPUs, GPUs or RAM to run your code on, without having to restart your instance, meaning rapid experimentation and flexibility for running your code against varying data and processing demands. 
User-Managed Notebooks are ideal for Data Scientists requiring more rigorous control of their environment. These notebooks are Deep Learning Virtual Machine (VM) instances, in which you can detail the machine type and framework for your VM instance. Notebooks can be configured with CPUs or GPUs, and a suite of pre-installed deep learning packages that supports TensorFlow, PyTorch, scikit-learn, NLTK and more. You can further customise your notebook by installing desired packages from within the instance itself, allowing for full control of your environment.
The client was particularly drawn to the customisability of the notebooks, and the ability to scale to their desired machine type and frameworks rapidly to cater for varying processing and data requirements, which heavily holds them back when using local machines. One of the major pain points of the client was the manual overhead involved with their current workflow, and the scheduled orchestration of notebook runs was incredibly attractive, meaning that model retraining and experimentation could be automated. Various integrations with BigQuery and Git were also appealing as these connectors assist with the reduction of development time.
No-Code and Low-Code Capabilities
Concerning Machine Learning modelling, the first workshops revolved around pre-trained models and off-the-shelf APIs that could reduce the manual workload for Data Scientists whilst being highly performant on generic use cases. As the client regularly utilises unstructured text data for their Machine Learning solutions, the <a href="https://cloud.google.com/natural-language/docs/quickstarts">Cloud Natural Language APIs</a> were explored, with particular emphasis on state-of-the-art content classification and various entity analysis capabilities. 
Having an advanced in-house Data Science research team, the client was eager to also learn how they could use Vertex AI for custom NLP use cases with use case and sector-specific content. The client was impressed with how straightforward creating a labelled Managed Dataset was for text classification, and with the rapid nature of creating an AutoML model without code. 
AutoML Natural Language provides a simple interface to train powerful Large Language Models (LLMs) without having to configure distributed training strategies or worry about machine types and other hardware considerations. With LLMs very much being the trend, having such a service that allows for rapid training and high model performance is becoming far more popular amongst forward-thinking businesses.
The client was also keen to get hands-on with the data they had successfully stored in BigQuery. The first port of call was to explore <a href="https://cloud.google.com/bigquery/docs/bqml-introduction#:~:text=BigQuery%20ML%20democratizes%20machine%20learning,The%20bq%20command%2Dline%20tool">BigQuery ML (BQML)</a>, where we could train models from the BigQuery UI using SQL. Together with the client, we trialled a myriad of classification algorithms that BQML accommodates on their pre-embedded text dataset. The client was impressed with the performance of BQML and how it rapidly provided a challenging baseline for their custom models to try and beat. 
Vertex Explainable AI integrates very easily with BQML to help provide a better understanding of model decision-making, both on a global and local scale. The client was impressed at how they could analyse the reasoning behind model predictions on an example-by-example level, which is becoming ever more important in the widescale adoption of AI.
We also explored hyperparameter tuning via the <a href="https://cloud.google.com/vertex-ai/docs/vizier/overview">Vertex AI Vizier</a> tuning algorithm to improve model performance. BQML is becoming an increasingly popular choice amongst clients due to the ever-expanding model options, and clever integrations such as <a href="https://cloud.google.com/vertex-ai/docs/training/hyperparameter-tuning-overview">Hyperparameter Tuning</a> and <a href="https://cloud.google.com/vertex-ai/docs/explainable-ai/overview">Explainable AI</a>, whilst also being a managed service. This means that Google takes care of the compute resources and the user does not need to worry about any hardware or machine type used in training.
Vertex AI Matching Engine
The client was particularly interested in vector similarity-matching, and being able to do so at low latency. Thus, we introduced them to <a href="https://cloud.google.com/vertex-ai/docs/matching-engine/overview">Vertex AI Matching Engine</a>, which is a high-scale low latency vector database which allows for embeddings to be queried against indexes with billions of embedding vectors to retrieve the most similar items within the database. The service is ideal for recommendation engines, text similarity, image search and ad targeting systems and it is the same technology that is used for <a href="https://cloud.google.com/blog/topics/developers-practitioners/find-anything-blazingly-fast-googles-vector-search-technology">YouTube and Google Search</a>. 
By deploying the client’s pre-embedded data to an Index and were able to showcase the rapid querying offering of Matching Engine by retrieving the closest embeddings at incredibly low latency. This unique offering of Vertex AI further added to the client’s confidence that Google Cloud was the right platform for them.
<h3>Conclusion</h3>
Using our MLOps Tooling Evaluation Framework, the client performed a like-for-like comparison between Google Cloud’s Vertex AI and AWS SageMaker and decided to move Machine Learning capabilities to Google Cloud. How Vertex AI could reduce key pain points such as manual overhead and the variety of powerful and flexible services at the fingertips of Data Scientists vs on-prem was a key factor in the decision-making. Now, the client is rolling out its first use case on Vertex AI, taking advantage of the MLOps sessions we delivered, utilising Vertex Pipelines, and our <a href="https://datatonic.com/insights/vertex-ai-mlops-turbo-templates/">MLOps Turbo Templates</a> to accelerate the productionisation of their Machine Learning models.
Datatonic is <a href="https://datatonic.com/insights/datatonic-google-cloud-specialization-partner-award-machine-learning/">Google Cloud’s Machine Learning Partner of the Year</a> with a wealth of experience developing and deploying impactful Machine Learning models and MLOps Platform builds. Need help developing an ML model, or deploying your Machine Learning models fast? 
Have a look at our <a href="https://datatonic.com/insights/vertex-ai-mlops-platform/">MLOps 101 webinar,</a> where our experts talk you through how to get started with Machine Learning at scale or <a href="https://datatonic.com/contact">get in touch</a> to discuss your ML or MLOps requirements!

Vertex AI Assessment: Choosing MLOps Tooling

Datatonic acquires Canadian-based Cloud Analytics Consultancy, Montreal Analytics, to accelerate Global Expansion 
17th April 
Datatonic is pleased to announce the acquisition of Montreal Analytics, a leading Canadian-based Cloud Analytics consultancy, and key partner to cloud data technologies, including: Google Cloud, Looker, dbt Labs, and Snowflake.  
The acquisition scales Datatonic’s footprint in the US and Canadian markets, and expands the group’s capabilities to serve global clients, through local, best-in-class delivery across EMEA and Americas markets.    
&nbsp;
About Montreal Analytics 
Founded in 2020, Montreal Analytics comprises a team of modern data experts, who hold over 100 professional certifications and accreditations in data engineering and analytics. 
The team has worked with leading and fast-growing businesses in North America, including the top digital music service worldwide, the largest Canadian retailer, and cutting-edge healthcare technology platforms, and is known for their agile delivery, best-in-class deployments, and proven business impact. 
As part of the acquisition, Montreal Analytics will become Datatonic Americas. 
Cyril Marques, CEO of Datatonic Americas, says: “Montreal Analytics was built focusing on the team and on the impact we deliver to our clients from cutting-edge data technologies. Two fundamentals shared with Datatonic whose footsteps we strive to follow. It’s an honor to now take those steps together as one formidable organization.”
&nbsp;
About Datatonic 
Montreal Analytics joins Datatonic, a leading Cloud Data + AI consultancy, recognised four times as the Google Cloud Partner of the Year, and is one of the first Looker partners in UKI and EMEA. 
Delivering over 500 Cloud Data + AI projects, Datatonic has driven measurable business impact for enterprises, such as: <a href="https://datatonic.com/case-studies/ai-booster-vodafone-supercharging-ai-ml-at-scale/">Vodafone</a>, <a href="https://datatonic.com/case-studies/bt-ai-accelerator-artificial-intelligence/">BT</a>, <a href="https://datatonic.com/case-studies/cloud-data-engineering-advanced-analytics-sky/">Sky</a>, <a href="https://datatonic.com/case-studies/reliable-predictive-models-mlops-delivery-hero/">Delivery Hero</a>, <a href="https://datatonic.com/case-studies/product-recommendations-mulberry/">Mulberry</a> and <a href="https://datatonic.com/case-studies/image-recognition-lush/">Lush</a>.  
Lennert Acke, COO of Datatonic, says: “Both Montreal Analytics and Datatonic have built strong reputations among clients and partners, in designing and deploying best-in-class solutions, through modern technologies, strategic thinking and enablement of their clients’ teams. We’re excited to share our experiences with and learn from our Canadian counterparts, which will no doubt strengthen Datatonic’s offering and our ability to serve our clients.” 
&nbsp;
Datatonic’s Global Expansion 
The acquisition strengthens and expands Datatonic’s ability to serve the US and Canadian markets, as well as provide global customers with local, best-in-class delivery across EMEA and Americas markets.  
Louis Decuypere, CEO of Datatonic, concludes: “Since our establishment, Datatonic’s mission has been to deliver purposeful impact to clients, with technology as the enabler to drive positive change. Having known Cyril and the team at Montreal Analytics for some years now, I’m excited that this talented team will be joining the Datatonic group. The acquisition of Montreal Analytics will allow us to advance our mission at a global scale.”  
&nbsp;
Want to find out more? Save your space by registering for our <a href="https://app.livestorm.co/datatonic/driving-analytics-adoption-with-admiral-google-cloud-and-datatonic?type=detailed">upcoming webinar</a>, where we’ll be covering how to increase adoption of your data + analytics tools at scale.

Datatonic acquires Montreal Analytics to accelerate Global Expansion 

Authors: <a href="https://www.linkedin.com/in/valentin-cojocaru-01844730/" target="_blank" rel="noreferrer noopener">Valentin Cojocaru, Data Science Lead</a>, <a href="https://www.linkedin.com/in/mgela/" target="_blank" rel="noreferrer noopener">Matthew Gela, Senior Data Scientist</a>, <a href="https://www.linkedin.com/in/christellexu/" target="_blank" rel="noreferrer noopener">Christelle Xu, Machine Learning Strategy Lead</a>
AI and Machine Learning have rapidly evolved to become an integral part of our daily lives, and are being increasingly integrated into business processes to unlock value. With that, there is an increased necessity to ensure AI is developed equitably, securely, transparently and accurately. In particular, the recent wave of releases of models such as Bard and DALL-E to the public has undoubtedly brought an increased level of opportunity, but also an unprecedented spotlight on the use and misuse of AI.
The UK Government’s <a href="https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach" target="_blank" rel="noreferrer noopener">new whitepaper on AI Regulation</a> illustrates the importance of integrating Responsible AI into your AI systems. In the age of Generative AI, Responsible AI is more important than ever, and it is crucial for businesses to understand what it actually is. In this first blog of our Responsible AI series, we’ll discuss the context for Responsible AI, why it matters and what it means to develop AI systems responsibly.
<h3 id="h-why-responsible-ai-matters">Why Responsible AI Matters</h3>
“Technology is neither good nor bad; nor is it neutral.” – Melvin Kranzberg
Our society has developed an unconscious trust in technology and engineering, largely focusing on the boundless upside of technological advancement, while at the same time trusting that these technologies are built safely and fairly; we don’t commute to work fearing our trains might suddenly crash. The same mindset has been applied in the field of AI, and over the last decade, many organisations have <a href="https://www.statista.com/statistics/941137/ai-investment-and-funding-worldwide/" target="_blank" rel="noreferrer noopener">rushed to create and invest in Machine Learning teams</a> that focus on delivering business value as quickly as possible, often misunderstanding the implications of ML systems that are built with outcomes, and not processes in mind. We don’t need to dig far to see <a href="https://incidentdatabase.ai/" target="_blank" rel="noreferrer noopener">examples of this</a>, including some very <a href="https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing" target="_blank" rel="noreferrer noopener">high-profile</a> ones.
The good news is that Machine Learning as both a topic and industry solution has become widely adopted, and now industries, governments and the public have started to understand the nuances and implications of building AI systems. Guidelines and principles have been made available by bodies such as the ICO, providing organisations with standards in which to develop AI systems.
Increasing Regulations
Governmental bodies have proposed, and in some cases already created, sweeping legislation to both control the misuse of AI and instil greater transparency in building and monitoring AI systems in production. Early preparation and adoption of the requirements of these policies should be incentivised and enforced sooner rather than later, especially for companies that have built a large number of ML applications in their portfolio and platforms. Understanding regulations, auditing their ML systems, creating blueprints and frameworks to create responsible ML applications by design and actively enforcing these practices takes time and effort.
For example, the EU is expected to pass the <a href="https://artificialintelligenceact.eu/" target="_blank" rel="noreferrer noopener">EU AI Act</a> in 2023, constraining the use of ‘high-risk’ ML systems such as some of the systems used in areas such as infrastructure, policing, biometric surveillance, and recruitment. Another key aspect of the legislation is the inclusion of certain requirements when building and monitoring these systems, such as the use of high-quality data, documentation, traceability, transparency, human oversight, accuracy and robustness.
In the US, federal, state and municipal laws already touch on ML applications, such as the Equal Credit Opportunity Act and the Fair Housing Act at a federal level, while at the state level New York, Illinois and Maryland have passed laws to mandate bias audits for automated employment decision tools that use AI to assist in candidate screening. More broadly, the White House has released the <a href="https://www.whitehouse.gov/ostp/ai-bill-of-rights/" target="_blank" rel="noreferrer noopener">Blueprint for an AI Bill of Rights</a>, which provides guidance for companies and individuals to protect them from threats presented by AI. China has also passed a regulation that controls the use of algorithms in online recommendation systems.
Developing AI Systems Can be Risky
Outside of any potential legal implications of not adhering to laws and regulations, there are other, equally important reasons why a company should build and maintain responsible ML systems. Companies should consider the cost of a wrong outcome or decision in their business. Could the system have an impact on the customer lifetime value? Are there any other brand or reputational risks involved? Could the system produce deep harm to end users or the general public? Are there any operational risks involved?
Responsible AI is most critical in systems that involve or affect decisions that impact human well-being, including access to resources or their health. For a company, Responsible AI can act as a differentiator, particularly with customers who choose to do business with organisations that align with their values.
<h3 id="h-what-does-it-mean-to-develop-ai-responsibly">What Does it Mean to Develop AI Responsibly?</h3>
The emerging field of ML risk mitigation practices has been coined as Responsible AI. Responsible AI does not have a strict and standardised definition, and we should instead aim to define it as a set of principles that govern the responsible and ethical use of AI, in a way that fairly impacts customers and society as a whole. It is usually comprised of the following five pillars:
Bias + Fairness
Fairness in Machine Learning means ensuring that the decisions and predictions made by a model should not unfairly advantage or disadvantage any particular individual or group in society. All ML models show some level of bias, but significant bias, or bias within high-impact use cases, can cause legal liability or real-world consequences. <a href="https://aclanthology.org/2021.nuse-1.5.pdf" target="_blank" rel="noreferrer noopener" data-lf-fd-inspected-kn9eq4rr9wy4rlvp="true">Systemic bias</a> and statistical bias can lead to differential outcomes and discrimination. This can have serious ramifications in domains such as criminal justice, hiring and lending.
One high-profile example of this is the criminal risk assessment algorithm (<a href="https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm" target="_blank" rel="noreferrer noopener">COMPAS</a>) used by judges and parole officers to predict the likelihood of reoffending in some US states, which was found to give significantly higher risk scores to people from ethnic minority backgrounds. Since data is tightly coupled to the functionality of ML models, care needs to be taken when collecting and understanding data, as well as selecting features for modelling. Detecting, mitigating and reporting bias in ML models across the modelling pipeline (from checking for data integrity at the time of collection all the way to reporting model results against feature distributions) is crucial.
<figure class="wp-block-table">
<table>
<tbody>
<tr>
<td>Note: In the blogs to come in this series, we will explore how to implement and monitor bias detection and mitigation methods within an ML production pipeline.</td>
</tr>
</tbody>
</table>
</figure>
&nbsp;
Explainability
Explainability in AI is the set of processes and methods that allows human users to understand and trust the results and output created by machine learning algorithms.
In order to determine whether your AI system is safe for making decisions or augmenting business processes, it is crucial to understand how it comes to a decision. Without this, you can have situations where the model could be basing its predictions on information it should not be relying on. Explainability enables developers to debug their models, stakeholders to understand whether the AI system is fit for purpose, and end users to qualify whether they can trust the model output.
Activities such as engineering meaningful features, generating explanations for individual model predictions, and setting up systems for tracking and creating thorough documentation of models are all ways to incorporate explainability into AI systems.
<figure class="wp-block-table">
<table>
<tbody>
<tr>
<td>Note: In the blogs to come in this series, we will explore the use of data cards and model cards as key documentation for improving the transparency of your models with key stakeholders, as well as feature-based and example-based explanations for understanding the decisions your models make.</td>
</tr>
</tbody>
</table>
</figure>
&nbsp;
Sustainability
Sustainability in AI is the practice of developing and using machine learning models in a way that minimises their environmental impact and promotes long-term societal and economic sustainability. As the use of AI and machine learning grows rapidly, the environmental impact of these technologies becomes more significant.
By adopting sustainable practices, we can ensure that these technologies are used in a way that benefits both society and the planet. Measuring energy consumption allows you to determine whether your energy usage is high, and provides insight into areas where consumption could be reduced. In a previous blog, we looked at how to predict the <a href="https://medium.com/datatonic/predicting-and-reducing-energy-consumption-of-machine-learning-models-7dc44e197225" target="_blank" rel="noreferrer noopener">energy consumption</a> of machine learning models based on the number of CPU operations they require.
Other methods to become more sustainable include reducing the amount of data required for training and deploying models, using models that require fewer resources to train and use, as well as optimising hardware usage by using compute resources more efficiently.
Robustness
Robustness ensures that machine learning systems can be monitored and perform reliably in different scenarios and over time. A robust AI system should be able to generalise well to new or unseen data and be resistant to minor changes in the data distribution. In many domains, such as healthcare and autonomous vehicles, the reliability of machine learning models is crucial. A model that is not robust may make incorrect decisions even when the input data is only slightly different from the training data, leading to undesirable consequences.
To improve robustness, you can set up monitoring and alerting designed to perform checks such as detecting data drift and sending an alert when the drift is high. Additionally, validating that the input data to a model is good quality allows the model to train and predict on clean and consistent datasets, which is a crucial step in improving the robustness of a model. In our Vertex AI Tips &amp; Tricks series, we have discussed how you can implement <a href="https://datatonic.com/insights/vertex-ai-data-validation-pipelines-great-expectations/" target="_blank" rel="noreferrer noopener">data validation</a> in your machine learning pipelines using Great Expectations, and how you can set up <a href="https://datatonic.com/insights/vertex-ai-setting-alerts-pipelines-google-cloud-monitoring/" target="_blank" rel="noreferrer noopener">alerting with Google Cloud Monitoring</a>.
Privacy + Security
Privacy and security in AI refers to the protection of sensitive and confidential data used by machine learning models, and ensuring that the models are secure against adversarial attacks.
Taking privacy-preserving measures is crucial for protecting the personal information that organisations store about individuals, preventing it from being misused or shared without consent. Fortunately, there are techniques that can be implemented to enhance privacy, such as data anonymisation, using privacy-preserving machine learning algorithms, or using a federated learning strategy to allow models to be trained across multiple devices or servers without exchanging data.
Model security prevents unauthorised access to machine learning models, protecting them from theft, modification or misuse. Models may also be susceptible to adversarial attacks, which is where attackers can generate examples that are designed to deceive the model into making incorrect predictions, leading to undesirable consequences. Techniques such as adversarial training and <a href="https://deepai.org/machine-learning-glossary-and-terms/defensive-distillation#:~:text=What%20is%20Defensive%20Distillation%3F,is%20less%20susceptible%20to%20exploitation." target="_blank" rel="noreferrer noopener">defensive distillation</a> can improve the model’s resistance to adversarial attacks.
<h3 id="h-conclusion">Conclusion</h3>
As AI becomes more widespread, it is crucial that we consider all five key pillars of Responsible AI. In the upcoming blogs in this series, we will expand on how your development teams can incorporate these pillars within your machine learning workflows, and we will demonstrate how to implement and monitor these within a production pipeline. We’ll also look at the wider picture, and give our take on why we believe Responsible AI is crucial in shaping the story of Generative AI towards something beneficial for society.
P.S. Although very tempting, this blog was not written by generative AI…
Datatonic is <a href="https://datatonic.com/insights/datatonic-google-cloud-specialization-partner-award-machine-learning/" target="_blank" rel="noreferrer noopener">Google Cloud’s Machine Learning Partner of the Year</a> with a wealth of experience developing and deploying impactful Machine Learning models and MLOps Platform builds. Need help developing an ML model, or deploying your Machine Learning models fast? 
Have a look at our <a href="https://datatonic.com/insights/vertex-ai-mlops-platform/" target="_blank" rel="noreferrer noopener">MLOps 101 webinar,</a> where our experts talk you through how to get started with Machine Learning at scale or <a href="https://datatonic.com/contact" target="_blank" rel="noreferrer noopener">get in touch</a> to discuss your ML or MLOps requirements!

Responsible AI 101: What You Need to Know

Authors: <a href="https://www.linkedin.com/in/kumail-kermalli-583540138/" target="_blank" rel="noreferrer noopener">Kumail Kermalli, DevOps Engineer</a> and <a href="https://www.linkedin.com/in/pooladkhay/" target="_blank" rel="noreferrer noopener">Mohammad Pooladkhay, DevOps Engineer</a>





<h5 class="wp-block-heading" id="3600">Background</h5>



<a href="https://cloud.google.com/build" target="_blank" rel="noreferrer noopener">Cloud Build</a> is Google Cloud’s serverless CI/CD platform. First launched in 2016, it allows users to execute builds in a fully-managed environment. Each build is made up of a series of steps and each runs in a specified Docker container. This allows a degree of customisation, as users can choose to use either <a href="https://github.com/GoogleCloudPlatform/cloud-builders" target="_blank" rel="noreferrer noopener">Pre-Built Images</a> or create their own to run their jobs on.



In 2021, an additional layer of customisation was added with the launch of <a href="https://cloud.google.com/build/docs/private-pools/private-pools-overview" target="_blank" rel="noreferrer noopener">“Private Pools”</a>. This allowed the setup of dedicated worker pools to execute a build, with the ability to access resources in a Virtual Private Cloud (VPC) network. Other useful customisation options include an increase in the number of concurrent builds allowed, greater machine type choice and the ability to disable access of your private pool to the public internet.



However, a key functionality that is not available in Cloud Build (at the time of writing, February 2023) is the ability to set a static external IP address to your build, even if it is running in a private pool. This would be extremely useful for some users as it would allow them to whitelist worker pool addresses in external providers they may need to access for builds.



Unfortunately, a Serverless VPC Connector and/or Cloud NAT cannot be used here due to private worker pools being located in a Google-owned Virtual Private Cloud network called the service producer network, peering into a project VPC (see Figure 1). Whilst Google does provide a <a href="https://support.google.com/a/answer/10026322?hl=en" target="_blank" rel="noreferrer noopener">list of external IP addresses</a> Google Cloud resources use, whitelisting all of them is impractical.
<h5>Introduction</h5>





Adding the native ability to assign an external IP address to Cloud Build is a <a href="https://issuetracker.google.com/issues/197128153" target="_blank" rel="noreferrer noopener">feature request</a> on Google’s Issue Tracker, so we may see this implemented in the near- to mid-term. However, in the interim, an alternative solution is needed. Read on to learn how to assign a static external IP address to a Cloud Build instance through the use of a proxy server hosted on Google Compute Engine.



The aim is to prevent Cloud Build instances from directly reaching out to the public internet over IP addresses we do not know the range of. Instead, a VM with a static IP address will act as their Proxy Server and all the communications between the VM and Build instances will be via private IP addresses.



<figure class="wp-block-image"><img decoding="async" src="https://cms.datatonic.com/app/uploads/2023/03/12Jjugu4UC4ZS1FdeINQsCQ.png" alt="Connecting our Cloud Build instance to an external static IP address" />
<figcaption class="wp-element-caption">Fig 1: Connecting our Cloud Build instance to an external static IP address</figcaption>
</figure>



<h5 class="wp-block-heading" id="6677">Tutorial</h5>



TLDR; Assuming you already have a VPC network and required IAM permissions:



<ol>
<li>Enable the Cloud Build and the Service Networking APIs</li>



<li>Create a VPC network (if you don’t already have one)</li>



<li>Allocate a named IP range in your VPC network</li>



<li>Create a private connection to the service producer network</li>



<li>Create a new Worker Pool</li>



<li>Create a subnetwork for the VM</li>



<li>Allocate a Static IP address to be added to the VM</li>



<li>Create a VM instance</li>



<li>Set up an HTTP Proxy Server on the VM</li>



<li>Add Firewall Rules allowing Ingress Traffic to the Proxy only via its private IP address</li>



<li>Instruct Cloud Build instances to use the VM’s private IP address as Proxy</li>
</ol>



If you wish to use the Google Cloud Console or gcloud CLI to achieve the goal, the process is well documented on both the <a href="https://cloud.google.com/build/docs/private-pools/set-up-private-pool-to-use-in-vpc-network" target="_blank" rel="noreferrer noopener">Networking</a> and <a href="https://cloud.google.com/build/docs/private-pools/create-manage-private-pools" target="_blank" rel="noreferrer noopener">Cloud Build</a> components as well as <a href="https://cloud.google.com/compute/docs/instances/create-start-instance" target="_blank" rel="noreferrer noopener">VM instances</a>. In this tutorial, we will demonstrate how to use Terraform to implement the above steps.



Source code is available at <a href="https://github.com/teamdatatonic/blog-cloud-build-static-ip" target="_blank" rel="noreferrer noopener">https://github.com/teamdatatonic/blog-cloud-build-static-ip</a> with comments indicating each step inside the main.tf file.



<ol>
<li>Enabling the Cloud Build and the Service Networking APIs</li>
</ol>



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">resource "google_project_service" "enable_cloud_build" { project = var.project_id service = "cloudbuild.googleapis.com" disable_dependent_services = true } resource "google_project_service" "enable_service_networking" { project = var.project_id service = "servicenetworking.googleapis.com" disable_dependent_services = true }</pre>





2. Creating a VPC network



Almost every resource in Google Cloud requires a VPC network to be added to.



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">resource "google_compute_network" "vpc_network" { project = var.project_id name = var.vpc_network_name auto_create_subnetworks = false mtu = 1460 }</pre>





3. Allocating a named IP range In your VPC network



We need to allocate a named private IP range so that we can later use it to peer our VPC network to the VPC network where Cloud Build Private Pools reside (Service Producer Network).



Cloud Build reserves the IP range 192.168.10.0/24 for the Docker bridge network. When allocating the IP ranges for resources in your project, Google recommends selecting a range outside of 192.168.10.0/24 in cases when Cloud Build builders access these resources.



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">resource "google_compute_global_address" "named_private_ip" { provider = google-beta name = var.named_private_ip_name project = var.project_id purpose = "VPC_PEERING" address_type = "INTERNAL" network = google_compute_network.vpc_network.name address = var.named_private_ip prefix_length = 24 }</pre>





4. Creating a private connection to the service producer network



This will create a private connection between our VPC network and Google’s own VPC network where our private worker pool will be hosted, i.e., the service producer network, using the private IP range we just allocated.



This step depends on the last step since we have referenced named_private_ip’s name.



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">resource "google_service_networking_connection" "service_producer_connection" { network = "projects/${var.project_id}/global/networks/${google_compute_network .vpc_network.name}" service = "servicenetworking.googleapis.com" reserved_peering_ranges = [google_compute_global_address.named_private_ip.name] }</pre>





5. Creating a new Worker Pool



When creating a worker pool, we want to prevent the Cloud Build instances from directly reaching out to the public internet and instead proxy all of their traffic through a Proxy server that we will host on a VM instance.



To achieve this, we set no_external_ip to true so that the instances won’t have direct access to the public internet. Then, we set the peered_network to our VPC network in which we created a private connection to the service producer network.



Finally, we need to make sure that the creation of the Worker Pool happens after the connection to the Service Producer Network by explicitly adding the depends_on.



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">resource "google_cloudbuild_worker_pool" "private_worker_pool" { name = var.private_worker_pool_name location = var.region project = var.project_id worker_config { disk_size_gb = 100 machine_type = "e2-medium" no_external_ip = true } network_config { peered_network = "projects/${var.project_id}/global/networks/${google_compute_network.vpc_ network.name}" } depends_on = [google_service_networking_connection.service_producer_connection] }</pre>





6. Creating a subnetwork for the VM



We need a subnetwork with a private range that we define so that later we can assign a private IP address from that range to the VM by which Build instances will communicate with the VM.



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">resource "google_compute_subnetwork" "proxy_subnet" { name = "${google_compute_network.vpc_network.name}-proxy-subnet" ip_cidr_range = var.vm_subnet_range network = google_compute_network.vpc_network.name region = var.region private_ip_google_access = false project = var.project_id }</pre>





7. Allocating a Static IP address to be added to the VM



This will be the public IP address for the Cloud Build instances.



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">resource "google_compute_address" "static_ip" { name = var.static_ip_name project = var.project_id region = var.region address_type = "EXTERNAL" network_tier = "STANDARD" }</pre>





8. Creating a VM instance



This VM will be our Proxy Server. Cloud Build instances talk to it via its private IP then it proxies their requests to the destination.



Keep in mind that VM’s specs should be adjusted to meet your requirements and var.vm_ip_address must be from the specified range in Step 5.



There is also a startup script added to the VM which we will discuss in the next step.



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">resource "google_compute_instance" "proxy_vm" { name = "proxy-vm" project = var.project_id zone = var.zone machine_type = "n1-standard-1" boot_disk { initialize_params { image = "debian-cloud/debian-11" } } tags = ["proxy-srv"] metadata_startup_script = file("setup_proxy.sh") network_interface { subnetwork = google_compute_subnetwork.proxy_subnet.name network_ip = var.vm_ip_address access_config { nat_ip = google_compute_address.static_ip.address network_tier = "STANDARD" } } }</pre>





9. Setting up an HTTP Proxy Server on the VM



You might have noticed that there is a startup script added to the VM. This script will run right after the VM gets created. It sets up a Forward Proxy on the VM so no manual work is required.



Feel free to use any Proxy Server that suits your needs. Here we are using “github.com/elazarl/goproxy” to build our own proxy server.



The script creates a Go project, builds it, and then runs the binary using a Systemd Service so it will start automatically on VM reboots.



You can change the Listen Port number for the proxy by editing the script. However, if you change it, you also need to change the firewall rule to match the new port.





10. Adding Firewall Rules allowing Ingress Traffic to the Proxy Port



For security reasons, we only accept ingress traffic to the Proxy Server from its Internal IP address. This means that no one from the outside world can use our Proxy Server.



This policy is enforced using a firewall rule:



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">resource "google_compute_firewall" "proxy_ingress" { name = "allow-proxy-ingress" network = google_compute_network.vpc_network.name source_ranges = ["${var.named_private_ip}/${var.named_private_ip_prefix_length }"] allow { protocol = "tcp" ports = ["9231"] # If you change this, make sure you're also changing it inside setup_ proxy.sh } target_tags = ["proxy-srv"] }</pre>





11. Instructing Cloud Build instances to use VMs private IP address as Proxy



Now that we have everything in place, it’s time to trigger a sample build to see the results.



Here is a simple Build instruction that sets the http_proxy and https_proxy environment variables, installs curl, and then queries ipify API. After a successful run, you should be able to see that the VM’s static IP address gets printed, indicating that the outside world thinks our Build Instances have that IP address.



<pre class="wp-block-preformatted has-background" style="background-color: #f1f1f1;">steps: - name: ubuntu script: | export http_proxy="http://10.0.3.10:9231" export https_proxy="http://10.0.3.10:9231" apt update apt install curl -y curl https://api.ipify.org/ options: pool: name: "projects/{project_id}/locations/europe-west2/workerPools/{private_worker_pool_name}"</pre>





<h5 class="wp-block-heading"> </h5>
<h5 id="d794">Conclusion</h5>



That is all! We have managed to assign a static external IP address to a Cloud Build instance through a proxy server hosted on Google Compute Engine. We have also verified this by running a sample Cloud Build job, where we saw that an external request&#8217;s IP address matched our proxy VM&#8217;s external IP. With a static external IP for your Cloud Build instance you can whitelist the IP in external providers you may need to access for builds, adding that extra layer of security to your application!



Datatonic, Google Cloud’s 4 X Partner of the Year, is the leading cloud data + AI consultancy for the world’s most ambitious businesses, challenging traditions to deliver tangible innovation at the leading edge of Google Cloud. Our team has a wealth of experience with Cloud Build, and other Google Cloud products. Need help with your data and AI? Get in touch at <a href="https://datatonic.com/contact/" target="_blank" rel="noreferrer noopener">datatonic.com/contact/</a>.

Configuring Cloud Build to use Static External IP Addresses

<h2>A guide by data leaders, for data leaders, looking to turn Data Mesh principles into practice on Google Cloud.</h2>
The evolving nature of technology and rapid growth of data volume, velocity, and variety presents a significant opportunity for businesses. However, companies often struggle to scale their more traditional data stacks. Data Mesh has gained momentum as an alternative, taking a decentralised approach to data management, and enabling businesses to harness and share data effectively to address challenges around scalability.
Considering that Data Mesh is as much an organisational challenge as a technical one &#8211; how do you implement it?
In this guide, Datatonic’s Lead Architect, <a href="https://www.linkedin.com/in/andrew-harding-60633b18/">Andy Harding</a>, and Data + Analytics Strategy Lead, <a href="https://www.linkedin.com/in/nevile-reid-6b77a384/">Nevile Reid</a>, provide tactical advice on how to take Data Mesh principles outlined by <a href="https://www.thoughtworks.com/what-we-do/data-and-ai/data-mesh/creating-business-value-with-data-mesh-whitepaper">Thoughtworks</a> and put them into practice across the organisation. You will learn:
<ol>
<li>The core principles and benefits of Data Mesh</li>
<li>Tactical advice on applying Data Mesh principles, including tool recommendations such as BigQuery, Data Catalog, and Analytics Hub</li>
<li>Four tips on driving organisational change to coordinate between data domains and promote a Data Mesh culture</li>
</ol>

Data Mesh on Google Cloud

Vertex AI Tips + Tricks: Data Validation in Vertex Pipelines using Great Expectations

MLOps is a set of best practices, guidelines and technologies designed to deploy and maintain ML models in production. According to a recent <a href="https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-top-trends-in-tech" target="_blank" rel="noreferrer noopener">McKinsey and Company report</a>, MLOps can shorten the production time frame for ML applications by 90% and greatly reduce the required development resources. ​



This year, Datatonic has worked on MLOps projects for clients such as <a href="https://datatonic.com/insights/bt-ai-accelerator-press-release/" target="_blank" rel="noreferrer noopener">BT</a>, <a href="https://datatonic.com/insights/ai-booster-vodafone-supercharging-ai-ml-scale/" target="_blank" rel="noreferrer noopener">Vodafone</a> and <a href="https://datatonic.com/case-studies/reliable-predictive-models-mlops-delivery-hero/" target="_blank" rel="noreferrer noopener">Delivery Hero</a> and experienced some of the biggest trends and challenges for companies looking to implement MLOps at scale, including data-centricity, real-time ML, and addressing skills gaps. 



In this article, Datatonic’s Director of MLOps, <a href="https://www.linkedin.com/in/jamie-c-b40722115/">Jamie Curtis</a>, tells you how this year’s trends have answered key MLOps challenges and shares our top MLOps predictions for 2023.
<h5>The Rise of Centralised MLOps Platforms</h5>







Some companies, especially large ones, are realising that they need centralised fundamental MLOps platforms.



“This year, we developed large-scale centralised end-to-end MLOps platforms for clients such as <a href="https://datatonic.com/insights/ai-booster-vodafone-supercharging-ai-ml-scale/" target="_blank" rel="noreferrer noopener">Vodafone</a> and <a href="https://datatonic.com/insights/bt-ai-accelerator-press-release/" target="_blank" rel="noreferrer noopener">BT</a> to enable them to scale Data Science across the business, and automate as much of the ML lifecycle as possible. This includes automation of deployment processes, training and retraining, serving, and monitoring; they want to be able to automate everything.” &#8211; says Jamie Curtis, MLOps Practice Lead, Datatonic



A primary driver for this is to serve Data Scientists as it allows them to focus on Data Science and automate processes which are inhibitive to that, like networking, security controls, access permissions, infrastructure management, and data engineering.
<h5>Companies are Rapidly Building MLOps Skillsets</h5>







In 2022, we’ve seen a large realisation of the need for skillsets and roles for people that can work within ML. Companies are realising that they need unique skill sets they don’t yet have, like those of ML Engineers, to achieve their business goals.



According to <a href="https://www.oreilly.com/radar/ai-adoption-in-the-enterprise-2022/" target="_blank" rel="noreferrer noopener">O’Reilly’s 2022 State of AI Adoption in the Enterprise</a>, only 26% of enterprise organizations have models in production, with respondents blaming lack of key skills as the biggest barrier to deployment.This gap between required skills and available talent is one of the reasons that <a href="https://datatonic.academy/" target="_blank" rel="noreferrer noopener">Datatonic Academy</a> was built. By providing customised instructor-led training courses, Datatonic Academy has successfully supported some of our largest customers to upskill their teams.
<h5>Significant MLOps Trends in 2023</h5>







Based on a wealth of experience within MLOps, delivery projects for companies such as BT, Vodafone and <a href="https://datatonic.com/case-studies/reliable-predictive-models-mlops-delivery-hero/" target="_blank" rel="noreferrer noopener">Delivery Hero</a>, Datatonic has solved common challenges for companies looking to implement MLOps. After noticing the most significant trends and challenges for MLOps in 2022, here are our top three predictions for 2023.





<ol>
<li>Increased focus on Data-Centricity</li>
</ol>



<a href="https://www.youtube.com/watch?v=TU6u_T-s68Y" target="_blank" rel="noreferrer noopener">Data-Centric AI</a> is shifting the focus from a model-centric to a data-centric approach in AI. Because an ML model is only as reliable as the data used in it, improving the quality of data improves outcomes. To maximise the value of their ML, businesses need to implement best-in-class data practices. However, most data platforms don’t take their companies’ data to a level of maturity advanced enough for their ML platforms.



This leaves a gap where extra effort is required to process the data before the ML models can be trained efficiently.



There are two main ways to address this gap:



<ul>
<li>Developing data engineering capability by extending the reach of end data products </li>



<li>Developing data preprocessing capabilities by providing more scalable or sophisticated methods for data cleansing, data validation, data splitting and feature engineering</li>
</ul>



Solving this gap will be a significant area of focus moving forward to simplify the process of implementing MLOps. 





<ol start="2">
<li>Reducing Entry Barriers for Data Scientists</li>
</ol>



Companies are already starting to make investments in reducing entry barriers for Data Scientists as much as possible. For Delivery Hero, we created a centralised MLOps platform to meet the needs of Data Scientists across the business with a focus on increased automation and reduced entry barriers for Data Scientists. This enabled them to experiment with ML models and then deploy those models much more easily and quickly.



We’ve noticed that there is quite a significant initial investment in upskilling that Data Scientists need to commit to before they can engage effectively with the MLOps platform that enables them to focus on Data Science. When scaling a platform across a large company, it’s optimistic to assume that everyone will commit to this learning curve and meet the required skill levels. Tools such as Command Line Interfaces (CLIs) can significantly reduce the amount of time and money spent on upskilling large teams, especially when used at scale. 





<ol start="3">
<li>More Real-Time ML</li>
</ol>



MLOps will also enable more real-time ML. As MLOps becomes more advanced, more companies will look to implement real-time ML. This is ML that learns in real time and constantly improves based on the data that it receives. Over time, this will enable models to become more accurate than ever before with less manual training required.



The MLOps infrastructure that enables real-time ML is becoming more common and more accessible. With MLOps becoming easier for Data Scientists, many more companies will seek to implement real-time ML for a range of applications.
<h5>All of This is Possible with Google’s Vertex AI…</h5>







Vertex AI offers a great range of solutions and a great depth of solutions. It provides complex aspects of MLOps as well as low-code and no-code options like <a href="https://cloud.google.com/vertex-ai/docs/start/automl-users" target="_blank" rel="noreferrer noopener">AutoML</a> and pre-trained APIs that are based on Google’s open datasets which are a lot more accessible. Other more niche tools and services only offer either breadth or depth, but not both. 



It also integrates well with other Google Cloud products such as the <a href="https://cloud.google.com/natural-language" target="_blank" rel="noreferrer noopener">NLP API</a>. Google focuses a lot on auxiliary and forward-thinking solutions such as explainable AI and responsible AI. They’re things that aren’t necessarily widely used or in demand right now, but they will be in the near future. 



Vertex AI benefits directly from and builds upon, Google&#8217;s world-leading research into state-of-the-art Machine Learning. It’s grounded in Google&#8217;s commitment to open source, providing greater portability, security, and reduced vendor lock-in.



Finally, it’s fully modular; you can use Vertex AI&#8217;s built-in solutions, but you can also build customised capabilities on top of Vertex AI using best-in-class technologies from open-source communities and partners (such as <a href="https://www.nvidia.com/en-gb/" target="_blank" rel="noreferrer noopener">NVIDIA</a>, <a href="https://labelbox.com" target="_blank" rel="noreferrer noopener">Labelbox</a>, etc.). As organisations are increasingly relying on their AI/ML capability for competitive advantage, the need for flexible and innovative ML platforms is greater than ever before.
<h5>Conclusion</h5>







MLOps is a practice with several clear advantages for businesses looking to generate real business value. 



Over the next year, we expect to see the following:



<ul>
<li>An increased focus on data quality</li>



<li>Lower entry barriers for Data Scientists</li>



<li>More real-time ML</li>
</ul>



In 2023, MLOps will only become even more powerful in pushing the boundaries of ML and Data Science forward. 



&nbsp;



Datatonic is <a href="https://datatonic.com/insights/datatonic-google-cloud-specialization-partner-award-machine-learning/" target="_blank" rel="noreferrer noopener">Google Cloud’s Machine Learning Partner of the Year</a> with a wealth of experience developing and deploying impactful Machine Learning models and MLOps Platform builds. Need help developing an ML model, or deploying your Machine Learning models fast? 



Have a look at our <a href="https://datatonic.com/insights/vertex-ai-mlops-platform/" target="_blank" rel="noreferrer noopener">MLOps 101 webinar,</a> where our experts talk you through how to get started with Machine Learning at scale or <a href="https://datatonic.com/contact" target="_blank" rel="noreferrer noopener">get in touch</a> to discuss your ML or MLOps requirements!

In this article, Datatonic’s Director of Machine Learning, <a href="https://www.linkedin.com/in/jamie-c-b40722115/">Jamie Curtis</a>, tells you how this year’s trends have answered key MLOps challenges and shares our top MLOps predictions for 2023.

MLOps in 2023

<figure class="wp-block-image size-large is-resized"><img decoding="async" loading="lazy" class="alignleft wp-image-11705" src="https://cms.datatonic.com/app/uploads/2023/03/Oliver-Webb.png" alt="data mesh" width="100" height="100" />
<figcaption class="wp-element-caption">Author: <a href="https://www.linkedin.com/in/oliverjameswebb/" target="_blank" rel="noreferrer noopener">Oliver Webb, Content Marketing Executive</a></figcaption>
</figure>



At our <a href="https://datatonic.com/insights/data-mesh-from-principles-to-practice/" target="_blank" rel="noreferrer noopener">Data Mesh: From Principles to Practice webinar</a>, Datatonic had the pleasure of welcoming data experts, <a href="https://www.linkedin.com/in/richardoastler/" target="_blank" rel="noreferrer noopener">Richard Oastler</a>, Head of Data Engineering at Dojo, <a href="https://www.linkedin.com/in/jopage2/" target="_blank" rel="noreferrer noopener">Jo Page</a>, Director of Data and Data Science at OVO, and <a href="https://www.linkedin.com/in/bipul-kumar-645995/" target="_blank" rel="noreferrer noopener">Bipul Kumar</a>, Head of Data and AI Practice, Google Cloud.



Along with Datatonic’s Lead Cloud Architect, <a href="https://www.linkedin.com/in/andrew-harding-60633b18/" target="_blank" rel="noreferrer noopener">Andrew Harding</a>, they discussed turning the concept of Data Mesh into reality.



We’ve seen for many years that organisations need data; they use it to make better business decisions, feed ML models, make predictions, and  innovate. Historically, we have met this demand with technology, and have seen it evolve through various centralised solutions to collect data across the enterprise and serve it to the business.
&nbsp;
Introducing Data Mesh 







While we have seen lots of evolution in the technology space, the mindset of centralised data teams has remained the same. Data Mesh is an analytical data architecture and operating model where data is treated as a product and owned by teams that most intimately know and consume the data. 



Data Mesh proposes a decentralised approach where individual departments build their own data capability within structured data domains. Each domain is responsible for building reusable and shareable high-quality data assets called data products. Domain teams apply their own expertise when building the products to benefit the entire organisation. 



The following four principles work together for the successful implementation of Data Mesh:



<ul>
<li>Data as a product</li>



<li>Domain ownership</li>



<li>Self-service data platform </li>



<li>Federated computational governance</li>
</ul>



Combined, these principles eliminate the issues brought about by other data structures, such as data silos and inconsistency. 



During our webinar, the audience gave us great insights into some of the common questions businesses might have when starting their Data Mesh journeys, ranging from the skills required for successful implementation to the challenges of data lineage and quality. Have a look at what our Data Leaders had to say.
&nbsp;
Q + A: Your Questions Answered







1. Google is always staying one step ahead of the curve concerning emerging trends in Data and AI. How is Google approaching the Data Mesh hype and what are some of the biggest challenges on the radar that we’ll see as companies transition?



Google: Data Mesh is a transformational concept and it&#8217;s challenging the status quo by attempting to solve a key challenge which has been around for years. The challenge is to improve the agility around the usage of data and hence increase or maximise the value of data. 



On the technology side, many of the building blocks are there, but the maturity needs to improve. The tooling is still evolving, especially in supporting newer concepts like federated centralised governance. Elsewhere, if you look at the organisational setup, many organisations don&#8217;t have the roles and personnel required for effective Data Mesh implementation. For example, finding a data product owner for every domain in the organisation who is considering Data Mesh is difficult. 



It will take some time for it to become a common practice but as the toolchain matures and the skill set improves we will see more success stories and the perception will change.



“Tooling maturity and organisational change are super important when it comes to getting started in Data Mesh.” &#8211; Andrew Harding, Lead Cloud Architect, Datatonic
&nbsp;





2. While Data Mesh is starting to develop as an evolution from centralised data platforms, users are asking whether there is still a place for a traditional Data Warehouse or whether Data Mesh acts as a replacement. Do you still see a need for a Data Warehouse?



Data Mesh offers a decentralised approach, but many companies are finding that this does not mean getting rid of their existing Data Warehouse, and are instead opting for a hybrid approach. 



OVO: For us, there is still a need for a central Data Warehouse. With some data, it’s important that it is centrally owned, managed, and curated by a central data team. 



However, the breadth of data in the warehouse must be monitored to make sure that businesses don&#8217;t just duplicate what is in the domains because that defeats the point of the Data Mesh. We are also working on making sure that data is easily-discoverable and interoperable across domains and this should diminish the need for how much data is in the warehouse.



Dojo: It&#8217;s important to highlight that moving to a Data Mesh does not mean ripping up everything that you&#8217;ve already put in place. We very much adopted the hybrid approach here at Dojo. With regards to challenges with maturity, we hope that products like <a href="https://cloud.google.com/analytics-hub" target="_blank" rel="noreferrer noopener">Analytics Hub</a> will simplify that and we can allow it to become much more isolated in the future but for us right now, that single project outputting your product from each domain is working well.



&nbsp;



3. There will be several companies that have transitioned incrementally over the years from on-premise infrastructure to centralised solutions such as a Data Warehouse, that will already have somewhat more advanced data capabilities. What advice would you give to businesses looking to migrate to Data Mesh from legacy architecture?



There are many companies at the forefront of data for which a Data Mesh is just one step further on their data journey. However, this does not mean that Data Mesh is inaccessible for companies still operating with legacy on-premise infrastructure. By implementing Data Mesh in stages and prioritising the biggest issues, it does not need to be an overwhelming process.



Google: With Data Mesh, there are four principles, and you don&#8217;t have to start by implementing all four. If a business is migrating from legacy data architecture, it should identify its main pain point and then identify which one of those four or maybe more than one of those four can solve it. For example, if the biggest pain point is agility and we identify that our infrastructure is not flexible, then tackling the self-serve infrastructure is a good place to start. 



The advice I would give is to pick one of those four pillars which will solve your biggest problem and then identify one or two use cases, and then think about building a Data Mesh from there. There will be a lot of learning depending on the organisation, but prioritising issues to solve will help businesses migrate efficiently.



&nbsp;



4. How do you address skills gaps within domain teams and what can you do if teams aren’t starting from a very technical background?



The biggest point to understand is the niche differences between roles. It is important to know exactly what you are asking of each role and ensure that your teams understand this too. Understanding the skills and targets of each role will help to build balanced and capable teams.



OVO: The Data Mesh principles assume that you have the required skills. We&#8217;ve been on a journey to deeply understand roles such as Software Engineer and Data Engineer and look at how the two might work together. Businesses need to make it very clear what their expectations are of each role to fill those skills gaps. 



Dojo: Different roles each add different skill sets to squads. If you know what each role is capable of, you can address areas where some aspects of data are not fully understood by forming multi-disciplinary teams within those domains as part of your strategy. You need to be flexible and understand what the skill requirements are for each area and resource it in that way.



&nbsp;



5. Data lineage gives greater visibility while simplifying the ability to trace errors back to the root cause in data analytics processes. With Data Mesh, the domain data teams become responsible for the quality and usability of data. So, how can businesses approach the problem of data lineage with Data Mesh?



Dojo: We struggled with this in the early stages because the tooling is limited. We took two strategies on the transformation side.



Firstly, we built our own service to help with data transformation. This is difficult because there&#8217;s inconsistent naming between domains, standards, and approaches to how your presentation layer is constructed. We&#8217;ve got to the point now where we have an MVP that&#8217;s running as part of our CI/CD process that does that final check before things are merged into production.



Secondly, the other approach is a people-based approach where we&#8217;ve been very clear with our analysts about what the requirements are and what can break the Data Mesh. In the long run, we want to remove as much of that tacit knowledge as possible and put it into the tooling so no one needs to think about it. We would be able to operate in the same manner as if we had a large centralised repository as we did at the start.



&nbsp;



6. Businesses have expressed some curiosity about Master Data Management (MDM) when using a decentralised approach to storing data. What are your thoughts regarding Master Data Management concerning Data Mesh? For example, is duplicate data okay?



Google: Master data management is important whether it&#8217;s in a Data Warehouse, Data Lake, or a newer Data Mesh. We can immediately see some challenges with MDM in Data Mesh because, by nature, it&#8217;s a central piece of information. Is it specific to a domain, should it remain central, or should it become its own domain? This will depend upon the organisation, but, more generally, the focus should be on providing and governing access to master data or reference data without duplicating it across every domain.



&nbsp;



7. While separating data by different domains sounds useful in theory, how would you approach a scenario where you’ve got a use case that comprises multiple domains or products?



Dojo: It&#8217;s really important during the early stages of rolling out Data Mesh that you define your domains in a very clear way with the rationale behind them. We&#8217;ve been through several evolutions and if you&#8217;re getting to the point where you cannot put something in one of those domains, you&#8217;ve probably done it wrong. We&#8217;ve had to re-architect several times to counteract that but it&#8217;s a learning journey. Ultimately by design, they should be able to function in isolation with limited restrictions from those other domains.



&nbsp;



8. With companies storing their data throughout the organisation, users are starting to ask whether a product catalogue is a key enabler for discoverability. To make products more discoverable, do the domain teams need to embrace some form of product catalogue? 



OVO: I certainly resonate with the need for a data catalogue which would have data products within it. We found you can scale, particularly in your Machine Learning data products, those benefits by exposing them across multiple domains. This is where the need arose for the customer intelligence hub. If you build a Machine Learning data product within one domain, you can very quickly scale that by just exposing it through other domains. For that to happen, you need to know that a data product exists which brings the need for a catalogue and, very importantly, a common language to categorise other products. 
&nbsp;
Summary







To be successful in Data Mesh, you need to be holistic in your approach by understanding it from the perspective of people, processes, and technology. When implemented effectively, Data Mesh enables best-in-class data management capabilities. 



There are new organisational structures with the creation of these domains and therefore new roles and skills will be needed within those domains to build out data capabilities. Tooling and maturity are also important. We need to let some tools mature and, in the short term, businesses may benefit from building some of their own capabilities to address those shortcomings.



“It will take some time for it to become a common practice but as the toolchain matures and the skill set improves we will see more success stories and the perception [of Data Mesh] will change.” &#8211; Bipul Kumar, Head of Data and AI Practice, Google Cloud UKI
&nbsp;
Getting Started







If you&#8217;re considering the Data Mesh for your business, there are a few key considerations to help you get started. 



<ol>
<li>Understand where your organisation is currently. </li>
</ol>



A recurring theme from our discussions was the need to understand where your organisation is at the moment, before trying to start Data Mesh. This will inform your starting point, and ensure you have the right level of planning and understanding.



<ol start="2">
<li>Apply the principle that suits your business needs first.</li>
</ol>



Once you understand your current paint points, a solid next step is to apply just one or some of the Data Mesh principles such as thinking of data as a product. In other cases, starting to create that self-service data platform might be a very logical first step. However, we must note that in the long-run, Data Mesh functions best when all four principles are applied.



<ol start="3">
<li>If Data Mesh seems too abstract, start with a Data Mesh MVP.</li>
</ol>



Businesses should start with an MVP if they are struggling with the abstract nature of Data Mesh. For example, build a single domain or a single use case to start making some of the principles of Data Mesh more real to your organisation and foster some buy-in and sponsorship for such a large transformational program.



The key thing is to reflect on where you are in your organisation today, think about your strategy, and come up with an approach that works for you.





Need help deciding if Data Mesh could work for you? <a href="https://datatonic.com/contact/" target="_blank" rel="noreferrer noopener">Get in touch</a> to find out more.

Events

Whitepapers

Insights

Datatonic and Intel Whitepaper: How to be Hyper-Relevant

Bring Personalised Experiences to Customers

Download
the guide

Related

Data Mesh on Google Cloud

Events

Whitepapers

Insights

Datatonic and Intel Whitepaper: How to be Hyper-Relevant

Bring Personalised Experiences to Customers

Download the guide

Related

Data Mesh on Google Cloud

Download
the guide