The Problem

Enterprise operations teams are buried in noisy telemetry — infrastructure metrics, application logs, and transaction data — with no scalable way to detect anomalies, forecast capacity, or diagnose root causes without hours of manual investigation.

What I Built

• Engineered 133 features across 6 groups (temporal, rolling stats, z-scores, lag, rate-of-change, cross-metric) and built an ensemble of Isolation Forest + PyTorch Autoencoder achieving 94.4% precision and 84.9% F1 on 17,280+ events per run
• Built a LightGBM capacity forecaster with 14.84% MAPE — a 35.7% improvement over the Prophet seasonal baseline — tracked across experiments with MLflow
• Integrated RAG-powered LLM reasoning: detected anomalies are matched against historical incidents in ChromaDB, then synthesized by Claude into structured root cause analyses with remediation steps, severity estimates, and resolution time forecasts

Impact

Automated end-to-end operational intelligence: from raw telemetry ingestion through anomaly detection, capacity forecasting, and LLM-generated root cause explanations — replacing manual incident triage with a production-grade ML pipeline.