I organized the Kestrel Home repository to support a low‑cost, high‑throughput logistics‑return‑prediction service. The layout groups an executive memo, detailed notes, model artifacts, scripts, FastAPI service code, source modules, tests, requirements, and deterministic validation evidence, all under clear top‑level directories.
Model performance is measured by ROC‑AUC (0.7728) and PR‑AUC (0.3949), and I evaluate the economic impact of interventions such as a proactive confirmation call priced at ₹45 per call with a probability threshold of 0.34.
My analysis shows a net monthly saving of +₹10,785 when processing roughly 700 orders with the proactive call strategy, while a dispatch‑hold policy would lose –₹43,722 per month due to a 12 % cancellation rate.
The deployment incurs effectively zero operating cost (₹0.00/month) because inference runs locally on CPU in under 2 ms per request.