Building ML Products, Not Just Models

There is no shortage of talented data scientists who can build a model that performs well on a test set. The harder problem, and the one that determines whether an AI initiative creates real value, is turning that model into a product that people actually use.

The Model Is the Easy Part

A well-tuned XGBoost model or a fine-tuned transformer is only one component of a production ML system. Around it, you need data pipelines that handle messy real-world inputs, monitoring systems that detect when performance degrades, and interfaces that present predictions in a way that non-technical users can act on.

Google’s 2015 paper on technical debt in machine learning systems illustrated this with a now-famous diagram: the actual ML code in a production system is a small box surrounded by a much larger system of configuration, data collection, feature extraction, monitoring, and serving infrastructure^[1]. The ML code is often estimated at roughly 5% of the total codebase. The remaining 95% is infrastructure.

Where Projects Stall

Most AI projects that fail do not fail because the model was inaccurate. They fail because of one of three reasons: the predictions were not integrated into an existing workflow, the data pipeline broke when upstream systems changed, or the team that built the model moved on before it was properly handed off. Gartner has consistently highlighted this pattern, noting that the majority of AI project failures stem from issues with data quality, integration, and organisational readiness rather than model performance^[2].

Our Approach

At Revolution Analytics, we build ML products end-to-end. That means we think about deployment, maintenance, and user experience from day one, not as an afterthought once the model accuracy looks good on a Jupyter notebook.

Our products like Amua and Nunua Sasa were built with this philosophy. The models behind them are important, but what makes them useful is the engineering that surrounds them: reliable data ingestion, clear user interfaces, and monitoring that alerts us before users notice a problem.

Fintech

Amua

A web-based data analytics platform for quickly analysing the financial data and behavior contained in M-Pesa and bank statements.

Practical Lessons

A few principles we have learned along the way. First, start with the decision the user needs to make, then work backwards to what the model needs to predict. Second, build the data pipeline before the model. If you cannot reliably get clean data in production, the model does not matter. Third, plan for model retraining from the start, because the world changes and your model’s assumptions will drift.

The organisations getting real value from AI are not the ones with the most sophisticated models. They are the ones that treat ML as a product discipline, not just a research exercise.