Projects

Running open-weight models locally via llama.cpp — from what fits on a 16GB MacBook Pro to quantizing a 72B on a 256GB Mac Studio, and measuring what the compression actually costs.

72B Q6_K vs f16: quality within noise (+0.0057 PPL), decode 2× faster (10.1 vs 5.1 t/s) · Qwen 4B at ~50 TPS on 16GB

• AI/ML
• Systems

AST-based static analysis for AI/ML library use in Python codebases, shipped end-to-end as a paid GitHub App — scanner engine, Cloudflare Worker + Container webhook pipeline, KV entitlement gate, and Lemon Squeezy billing.

• AI/ML
• Tools

A closed-loop debugging system for the proxy platform — YOLO visual + functional QA gate, blue-green deploys with instant rollback, LLM + RAG on the novel-fix path, event-driven staleness tracking so learned fixes don't rot, and a Telegram approval loop for anything complex.

• AI/ML
• Systems

A stateless fleet health + load-balancing service for the proxy platform — weighted random selection with latency-based weights, hysteresis-protected health states, CAS-safe updates against Couchbase.

• Backend
• Systems

A multi-layer reverse proxy for upstream-resilient routing — OpenResty, Couchbase, and a TypeScript control plane. Serves ~10M requests/day in production from a 2 vCPU / 2 GB node.

10M req/day in production · 2 vCPU GCP e2-small · 60k concurrent at 100% in load test

• Backend
• Systems

A modular RAG framework for benchmarking retrieval strategies — vector vs. keyword, with and without rerankers — before anything touches an LLM.

• AI/ML

Transfer-learning YOLO on small, custom datasets for objects not in COCO — working the dataset quality and overfitting problems more than the model architecture.

• AI/ML