From PhD Research to Production ML: Lessons from the Lab

Academic ML and production ML optimise for different things. After publishing research at A-rank international computer science conferences and Q1 journals — while simultaneously shipping production systems — we've accumulated some concrete lessons on bridging that gap.

Where Research and Production Diverge

In research, you optimise for novelty and benchmark performance. In production, you optimise for reliability, maintainability, and business impact. These are different disciplines, and the lessons below reflect that.

Reproducibility matters more than novelty

In a paper, you need a novel contribution. In production, you need a model that gives the same results every time, on every deployment, with clear documentation of how it was trained and why decisions were made.

Simple models often win

The most sophisticated model isn't always the best choice. We've deployed gradient boosted trees that outperform deep learning approaches in production — they're more robust, faster to retrain, and easier for the client's team to understand and maintain.

Feature engineering is the real work

In both research and production, the quality of your features determines the quality of your model. The difference is that in production, you also need to think about feature availability at inference time, data freshness, and pipeline reliability.

Monitoring is not optional

A model that worked great at deployment can degrade over time as data distributions shift. We build monitoring and alerting into every production system from day one.

Putting It Together

A research background helps identify which techniques are worth applying to a given problem. Production experience determines whether they'll hold up in practice. The useful work happens at the intersection — picking the right approach and making sure it runs reliably in the real world.