ML-Powered Predictive Analytics Engine
A production ML system that predicts expected loss and claim severity using gradient-boosted ensembles trained on multi-year policy and claims data, deployed as a low-latency scoring service.
Gradient boosting
Algorithm
Real-time API
Inference
AutoML + MLflow
Ops
Problem & use case
Legacy actuarial models lacked the granularity to price emerging risks and were slow to retrain when market conditions shifted.
Solution & engineering approach
An automated ML pipeline with feature stores, model registry, and a FastAPI scoring layer — supporting champion/challenger experiments and real-time inference at scale.
Architecture & data flow
- 01
Feature engineering from 200+ policy, claims, and exposure variables
- 02
AutoML sweep across XGBoost / LightGBM / CatBoost
- 03
MLflow tracks experiments, artifacts, and lineage
- 04
Champion model deployed via FastAPI behind a load balancer
- 05
Drift monitoring + scheduled retraining loop
Technology stack
Implementation status
Implemented — ML pipeline built with MLflow model registry and FastAPI scoring layer