Machine learning stopped being a slide deck and started making real decisions — sales, marketing and risk models live across fintech, with the operating habits that made model number two cheaper than model number one.
In 2017 we put machine learning into production — not as an experiment, but as part of how a fintech business made everyday decisions. Models went live across sales, marketing and risk, including the credit scoring that sits at the heart of a lending platform.
The hard part of ML has never really been the first model. It's the second, the third, and the version you have to explain to an auditor two years later. That's the problem we set out to solve from the start.
We treated models the way we treat software: versioned, tested, monitored, and owned over time. The result was an operating discipline where model number two was cheaper to build than model number one, because the platform, the data pipelines and the review habits were already in place.
That discipline is the reason a lending client could run generation after generation of scorecards without ripping the system out and starting again — and the reason the risk numbers held up under scrutiny.
2017 is where our ML & AI Ops practice was really born. The same habits — retrain on schedule, roll back on a button, keep the audit trail — now run under AIOps in our clients' operations rooms and under the AI tooling we build for ourselves. The lesson from 2017 still holds: getting a model to production is easy; keeping a fleet of them honest is the actual job.
The year transformation became repeatable, not bespoke.
Read the story →