I lead engineering on OxCode, an AI software engineer for VS Code and the terminal. Version 0.9.0, the public launch, is live on the Visual Studio Marketplace and Open VSX.
My part spans the agent loop, the API gateway behind it, and the evaluation work that measures both. The run on the right is the real extension panel, recorded from its own webview: it finds a seconds versus milliseconds bug, fixes it, asks before running the tests, and stops when they pass.
Before OxCode I built the backend of Oxlo.ai, a serverless inference platform serving 45+ models across Azure AI Foundry, AWS Bedrock and our own GPUs.
Both runs are published with their predictions, reports and method in oxcode-evals.
- SWE-bench Verified: 410 of 500 (82.0%), with the agent offline inside each issue's own environment. See the run
- Terminal-Bench 2.1: 73 of 89, our best round. See the run
- Rate limiting in Redis, atomic per minute, hour and day, with concurrency limits
- Priority queues per model, ordered by plan, so a burst waits instead of failing
- GPU load balancing and autoscaling across T4 to H100 workers
- Caching, from a Redis user cache on the hot path to prompt caching that took Claude calls from zero cache hits to a cached prompt
- Retries and fallback routing across model providers in the OxCode gateway
Person Detection, Tracking and Re-Identification YOLOv8 detection, Deep SORT tracking and OSNet re-identification. Keeps a stable ID for each person through occlusion and re-entry.
AI Code Debugger A LangGraph workflow with parser, fixer and reviewer roles, iterative reasoning loops and model selection by cost.
Universal Document Intelligence Chatbot Answers across uploaded documents and live web search, routing each question to the source that can answer it, with citations.
PolicyPulse Scans documents and email for policy violations with configurable rules for personal data and confidential material.
- Languages: Python, TypeScript, JavaScript, SQL, Bash
- AI: coding agents, tool calling, RAG, prompt caching, LLM evaluation, LangGraph, PyTorch
- Models: DeepSeek, Kimi, Qwen, GLM, Llama, Mistral, Claude and GPT, plus FLUX, Whisper and YOLO
- Backend and infrastructure: FastAPI, Postgres, Redis, Docker Swarm, Azure, AWS Bedrock, RunPod


