Projects across deep learning, generative AI, and full-stack development.
Vigil is an autonomous, serverless MLOps pipeline on AWS for aircraft engine predictive maintenance. It predicts Remaining Useful Life (RUL) using an XGBoost model, actively monitors for data drift, generates LLM-powered incident reports, and seamlessly orchestrates model retraining and promotion without human intervention.
- Developed Stellwerk, a Rust-based LLM inference gateway (Axum/Tokio) with Redis-backed semantic caching and token-bucket rate limiting, instrumented end-to-end with Prometheus metrics - Load-tested the gateway under realistic concurrent traffic — 100 simulated users, ~300 requests — which surfaced an unbounded-queueing bug causing 43.2% request timeouts. - Resolved it by adding a bounded concurrency limiter that fails fast under saturation instead of queueing — cut timeouts to 0.7% (~62x reduction) and boosted throughput 4x (0.99 → 4.12 req/sec)