Signal over noise.

Three tables — labels, demographics, two years of transactions — joined into 970,960 subscribers at a 9% churn rate, described by ~50 engineered behavioral and tenure features.

Then the audit: every time-based feature checked against a strict 2017-03-31 cutoff — catching a date-parsing flaw that leaked future tenure as “active status”. Cleaned, LightGBM reaches 0.956 PR AUC, and the top 10% risk slice holds 94.8% of churners.

Every complaint is a signal; the problem is volume. TF-IDF compressed into a TruncatedSVD semantic space, a classifier with calibrated probabilities, and cosine-similarity retrieval matching historical cases in ~0.18ms — no GPU required.

Serving is engineered, not implied: a FastAPI REST server, a Streamlit interface, and SQL analysis of complaint patterns and dispute rates — plus an anomaly radar that notices when the shape of risk changes.

XGBoost and an Attention LSTM, each trained on 15-day sequential price action for NSE tickers and cached for high-throughput inference.

Gemini 2.5 Flash synthesizes macro drivers — RBI repo rates, Union Budget, tariffs — blended at a deliberate 60/40. Every inference and macro summary is persisted for historical lookup.

End-to-end architecture at Krea — modular workflows, business logic, REST APIs — substantial parts built with AI-assisted development in Claude Code.

The measure is operational: cafeteria peak waits fell ~30% after launch. React on the front, orchestration underneath, real users throughout — still running, still mine.