All Posts
March 10, 2026
From PhD Research to Production ML: Lessons from the Lab
What we learned translating academic ML research into production systems for real businesses.
March 10, 2026
What we learned translating academic ML research into production systems for real businesses.
There's a running joke in machine learning: academic papers show 99% accuracy, but production systems are happy with 80% that actually works. After years of publishing research at venues like GECCO, EvoStar, and IEEE TETCI — and simultaneously building production ML for businesses — we've learned a few things about bridging that gap.
Academic ML and production ML are different disciplines. In research, you optimise for novelty and benchmark performance. In production, you optimise for reliability, maintainability, and business impact.
Here are the biggest lessons we've learned:
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.
The most sophisticated model isn't always the best model. We've deployed gradient boosted trees that outperform deep learning approaches — not because they're more powerful, but because they're more robust, faster to retrain, and easier for the client's team to understand and maintain.
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.
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.
Our research background gives us a deep understanding of what's possible with ML. Our production experience tells us what's practical. The intersection of those two things is where we create the most value for our clients.
When we assess a business problem, we're drawing on published research to identify the right approach — and production experience to make sure it actually works in the real world.