WhereItLands
Decoupled ML Prediction Service, Stacking Engine & Dockerized CI/CD
- Engineered a decoupled, end-to-end ML service predicting outcomes with 64.24% out-of-sample accuracy, evaluated via 5-split TimeSeriesSplit to prevent temporal data leakage.
- Developed an asynchronous FastAPI REST microservice leveraging startup lifespan caching, delivering predictions with a median latency of < 8 ms and a compact ~76 MB memory footprint.
- Implemented a Stacking Ensemble (XGBoost, LightGBM, ExtraTrees, RidgeCV) coupled with a bivariate Poisson and Coles-Dixon statistical engine for exact-score density estimation.
- Automated testing and deployment workflows using GitHub Actions CI and Docker, guaranteeing pipeline repeatability across unit, integration, and endpoint suites.
PIPELINE OVERVIEW
v2.4 [PROD]
Validation Engine
5-Split TimeSeries
Stacking Ensemble
64.24% Accuracy
Inference Speed
< 8ms Latency
CI/CD & Runtime
Docker + GitHub Actions
