X-Ray

Embat's challenge asked whether money alone can tell you how a company is doing. Companies don't fail on the balance sheet, they fail on cash: annual ratios arrive late, while the daily trail of receipts, payments, and debt arrives on time. The catch was that the data had no default label to learn from.

I led the machine learning side: the forecasting lab (Huber and Ridge regressions validated with GroupKFold, MAE at 1, 3, and 6 months) and the structural forecast that projects each account forward and re-scores it with the same function. I also worked on the score itself and took part in the product and UI decisions.

A deterministic, point-in-time score over 24 months of banking data for 1,286 companies in 250 corporate groups. On top of it: a six-state trajectory machine, an additive waterfall that explains every point of change, a what-if simulator, group consolidation, and an alert monitor, all driven by the same scoring function.

• No default label, so no GBDT at the core: a rule-based, causal scoring function, with tests proving that future data never changes past scores.
• One function powers the score, the simulator, the outlook, and the structural forecast, so a simulation can never disagree with the score. ML lives next to it, not instead of it.
• Python and DuckDB engine behind a FastAPI backend, and a Next.js front end with a portfolio view for Embat and a 360° view for each company.
• Validated on synthetic banks with controlled regimes: 95.5% of deteriorations detected within six months, with zero structural false alarms on stable companies.

• Answers the six questions of the challenge: who's healthy, who's improving, who's slipping, dip or decline, why it changed, and how early it was visible.
• Deployed demo and a presentation video, with the limits of the model stated openly instead of hidden.

• Python
• FastAPI
• DuckDB
• scikit-learn
• Next.js
• TypeScript