I lead engineering on OxCode, an agentic AI software engineer that ships as a VS Code extension and a CLI. I own its orchestration core: the lifecycle in the header above, running on an explicit state machine, with per-role model routing, sub-agent dispatch, and a verifier designed to catch vacuous or weakened tests rather than trust the model's own report.
Before that I architected the Oxlo.ai backend from scratch: a serverless LLM inference platform with an OpenAI-compatible API, intelligent routing across Azure AI Foundry, AWS Bedrock and self-hosted GPU workers over 45+ models, quota-aware fallback chains, per-plan rate limiting, usage metering and billing. I deployed and auto-scaled LLM and diffusion inference on T4, L40S, A100 and H100 clusters.
Most of the difference between a coding agent that works and one that wastes your afternoon is not the model. It is the harness around it: context management, tool design, the verification loop, and whether the system tells you the truth about what it actually did. That is the problem I spend my time on, and it is why I benchmark instead of demo.
| OxCode | An AI software engineer for VS Code and the terminal. Published on the Visual Studio Marketplace. |
| Oxlo.ai | Serverless LLM inference platform. I built and own the backend. |
| oxcode-evals | Reproducible agent evaluations with published methodology and prompts. |
Person Detection, Tracking and Re-Identification YOLOv8 detection, Deep SORT tracking and OSNet re-identification. Holds a stable ID per person through occlusion and re-entry into frame.
AI Code Debugger A role-based LangGraph workflow (Parser, Fixer, Reviewer) with iterative reasoning loops and dynamic model selection for cost and performance tradeoffs.
Universal Document Intelligence Chatbot Answers questions across uploaded documents (PDF, DOCX, TXT, CSV, XLSX) and live web search, routing each query to whichever source can actually answer it, with citations.
PolicyPulse Document and email scanning that flags policy violations through configurable rules for PII, confidential material and compliance terms, with a dashboard and REST API.
Traffic Sign Detection Fine-tuned YOLOv5 for traffic sign recognition.
Stack as text
- Languages — Python, TypeScript, SQL, Bash, C/C++
- Agents and LLM — agent orchestration, tool calling, context management, RAG, evaluation harnesses, LangGraph, LangChain, ONNX embeddings
- Backend — FastAPI, Postgres, Redis, SQLAlchemy, Alembic, Next.js
- Infrastructure — Docker and Swarm, Azure (VM, Blob, Postgres, AI Foundry), AWS Bedrock, RunPod, SaladCloud, GPU serving on T4 / L40S / A100 / H100
- ML — PyTorch, Transformers, ONNX Runtime, OpenCV, YOLO, spaCy