03 / Flagship Engineering
Shipped
Lead Scoring ML API
An evaluated supervised-ML inference service with calibration, threshold policy, model integrity checks, and reproducible synthetic benchmarks.
FastAPIscikit-learnPandasPostgreSQLDockerPytest
Case study
What this project demonstrates
The project focuses on the engineering around an ML model: data contracts, probability calibration, decision thresholds, artifact integrity, reproducibility, and honest evaluation boundaries.
Architecture
Synthetic dataset → feature pipeline → model selection → calibration + threshold policy → signed artifact manifest → FastAPI inference
Engineering highlights
- Deterministic 5,000-row synthetic B2B dataset with explicit non-production benchmark labeling.
- Train/validation/test separation, calibration analysis, Brier/log-loss evaluation, and business-cost threshold policy.
- Single, batch, and CSV scoring with typed feature contracts and local logistic contribution explanations.
- Model manifest and SHA-256 verification before deserialization, CI coverage, CodeQL, and release/container workflows.
Boundaries / limitations
All benchmark metrics are synthetic and must not be interpreted as real CRM performance or causal business impact.