Unlocking Success With MLOPS: A Framework for Effective Adoption
The adoption of AI/ML is not just about leveraging technology; it’s about fostering an environment that allows data science to flourish, ensuring the business can reap the substantial rewards that come from successfully implemented machine learning.
We often hear about companies with armies of Data Scientists, Data Analysts and Researchers who fall short of delivering tangible value.
As a key player in the implementation of a formal AI/ML Ops practice at SSENSE, and through consulting work with diverse companies and non-profit organizations, I have gained deep hands-on experience in guiding and witnessing various iterations of this transformative process.
In this article, I will share key insights and personal notes from my experiences to facilitate the adoption of Machine Learning Engineering (MLEng) and Operations (MLOps) for tangible results.
Disclaimer
Data Science is notorious for its wide range of role definitions. A data scientist in one company might be called a data analyst in another, and a machine learning engineer in another. At SSENSE, data scientists work on modeling problems and finding solutions, sometimes theoretical, for them. Machine learning engineers are responsible for implementing these solutions and the necessary infrastructure for those solutions to operate at scale. This is the framework I’ll be using as a reference.
1. The right tool for the job
Any ML adoption should begin with a conversation with your data scientists. Their comfort and productivity with specific tools are crucial.
10 to 12 years ago, SAS or SPSS was the de facto standard, and it was costly. New talent lacked experience with these products so it made sense, for instance, to migrate from SAS to R.
Modern AI/ML has progressed with new languages and libraries are being added regularly.
Your focus should be on standardizing the execution of the code/tooling. This can be achieved by adopting the principle of “convention over configuration”. Provide containers and templates for a standard foundation so that both scientists and engineers feel at home.
However, this freedom comes at a cost. Choosing your tools does not mean neglecting the need for high code standards and comprehensive documentation. Remember, the code that makes it to production must be supported, so a clear understanding of what’s used is essential.
👉 Favour convention over documentation
👉 Provide containers and templates so that you can provide some standard foundations
👉 This can be language agnostic
👉 This will also reduce the high costs of having different sandbox environments for each individual and project
To keep on your radar:
👉 Whatever code gets created and makes it to production will need to be supported
👉 Although you may not want to enforce which tooling is used, you will need to be aware of what will be used
👉 Great documentation and high code standards are always essential
2. 0 to 100 real quick
Data scientists are valuable assets. To effectively harness their skills, it’s crucial to provide an environment that maximizes their potential to contribute to business value. This involves minimizing the time they spend on data development and infrastructure setup.
Aim for a “coding on the first day” environment for new hires, with consistently prepared and easily accessible data sets. Prevent the need for repetitive tasks with self-serve tools, ensuring that data sets used across teams are consistent.
This approach will also aid in staff augmentation should you need to scale rapidly.
👉 Provide curated data sets and a straightforward development environment
👉 Minimize the need to perform repetitive tasks with self-service tools
Take a look at how SSENSE tackles this by leveraging Data Mesh.
3. Highway to prod
One of the most significant challenges I’ve observed is the friction encountered on the path to production. Many promising ideas and innovations never make it due to issues such as data governance, infrastructure, testing, and conflicts with production releases. This friction often results from the data team working in isolation, and resorting to workarounds that don’t cut it in the production environment.
Regardless of your scientists’ skill level, this friction could be your biggest pitfall, affecting business revenue, team morale, and the credibility of the data team.
Plan ahead for releases, integrating data science team releases with your Business as Usual (BAU) activities. Encourage your ML team to learn from their software engineering counterparts’ delivery approaches.
Think ahead. You’ll be releasing a service — often for the first time — so plan for it in terms of business process, technology, and team culture.
👉 Integrate your Data Science team’s releases with your BAU
👉 Study how your counterparts approach delivery
👉 Tackle small in order to fail quickly and learn fast
4. Emphasizing AI/ML ethics
As AI/ML becomes more embedded in our society, the need to demonstrate ethical AI/ML practices intensifies. The “right to explanation” clause in GDPR allows consumers to challenge any decision made based on an algorithm that impacts them.
In my opinion, there is still no silver-bullet solution to this. I’ve seen companies that have attempted to tackle the issue at scale, they’ve all tended to implement strict, imprecise, and overly broad policies. This approach often leads to false positives in risk identification and hampers production. These problems hit the fan by orders of magnitude when you introduce third-party vendors, who may not be considering these issues at all.
At the very least, you should keep this issue on your radar. Proactively stay informed on the matter, keep an eye out for new technologies that could address the problem, and be aware of any upcoming regulations.
Do your due diligence to avoid using data sets or algorithms with inherent bias or proprietary paywalls. Ensure that your entire practice is aligned on these issues.
👉 Do your due diligence and stay up to date on AI/ML regulation
👉 Avoid using data sets or algorithms that have been trained with bias or a proprietary paywall
👉 Regulations are coming, plan ahead
To quote Warren Buffet: “Only when the tide goes out do you discover who’s been swimming naked.”
5. The nice way to ask
Clearly documented business and technical requirements are a fundamental pillar of successful ML adoption.
Instead of attempting to replicate working prototypes built by data scientists, focus on business requirements tied to company objectives and any constraints that may exist.
Remember, data science is a science. Empower your team to do their job, provide requirements tied to business objectives, and let them discover the best solutions.
From an ROI perspective, I’ve found it much more efficient to provide clear directives to data teams rather than posing open-ended questions. The output will be more tangible and digestible, but also provide long-term value. Answering questions should be part of the process, not the end goal.
Good
👉 Forecast customer lifetime value
👉 Detect fraudulent transactions
👉 Reduce overbuys
👉 Optimize this or that
Bad
👉 Who are our best customers?
👉 How can we optimize our purchases?
👉 Should we implement this or that?
6. AI/ML and business sitting in a tree
Embed your data team into your business, or vice versa, to foster a more comprehensive understanding of your organization’s challenges. Even better, if resources allow, appoint a Subject Matter Expert (SME) to support the AI/ML team full-time, aligning their Key Performance Indicators (KPIs) with the value derived from these efforts.
This approach will significantly enhance your ROI and accelerate results.
Often, the most innovative ideas and insights come from thinking outside the box. Data science teams can provide a broad overview of business objectives while ML teams are your best allies to know what can and can’t be brought to life. With deeper integration, they can gain a more detailed view of each department.
This integration enables them to identify previously unseen problems and work to optimize or provide solutions.
Make sure your team is close to the business by design.
👉 Witness the challenges SMEs face each day
👉 Empower your team to identify issues that they can solve
👉 Sometimes the business is too close to see the problem
👉 All domains should learn and benefit from the successes of other departments
Conclusion
Adopting AI/ML is a challenging yet rewarding journey.
By focusing on empowering your data scientists, establishing a conducive environment, and creating a frictionless path to production you will ensure your team can focus on solving value-added problems. By prioritizing clear business requirements, and integrating your data science team into your business, you can facilitate a successful transition.
These simple practices helped me drive significant change, and I am confident they can help your organization unlock the full potential of AI/ML. Remember, MLEng and MLOps is not just a technological shift; it’s a paradigm shift from finding viable solutions to effectively delivering them.
Editorial reviews by Catherine Heim & Mario Bittencourt
Want to work with us? Click here to see all open positions at SSENSE!
