AI agent engineer with production C++ experience.
I build coding-agent infrastructure, repository intelligence systems, and deployable AI workflows with a bias toward reliability, auditability, and deterministic verification.
Most repositories here were recently polished and open-sourced from local/private development work. I optimize for runnable demos, tests, explicit limitations, and reproducible engineering evidence rather than commit-count signaling.
- Coding-agent safety: sandboxed execution, policy checks, audit trails, replayable artifacts
- Repository intelligence: code-aware retrieval, citations, MCP tools, approval-gated maintenance
- Local-first agent design: deterministic pipelines, BM25/AST retrieval, test-driven verification
| Project | What it demonstrates | Status |
|---|---|---|
| RepoAirlock | Safety-oriented execution harness for coding agents: Docker isolation, policy enforcement, audit trails, patch replay, HTML reports | v0.1.1-alpha |
| CodePipe | Deterministic local-first coding agent with a fixed expert pipeline, BM25 + AST retrieval, fuzzy patching, and pytest verification | v0.1.0-alpha |
| RepoRAG | Source-grounded repository RAG plus an approval-gated code maintenance workbench with LangGraph, MCP, and security guards | v0.1.0-alpha |
| GEO Lens | Production-style full-stack AI SaaS for GEO / AI-search visibility workflows | Live Demo |
| Layer | Project | Purpose |
|---|---|---|
| Repository understanding | RepoRAG | Retrieve code context, answer with citations, expose MCP tools |
| Deterministic code editing | CodePipe | Local-first repair pipeline with BM25/AST localization and pytest verification |
| Safe execution | RepoAirlock | Run coding agents in Docker with policy checks, audit logs, and replayable reports |
| Productization | GEO Lens | Full-stack AI SaaS demo for AI-search visibility workflows |
| Vision / edge AI | DMS-MultiTask | Lightweight driver-monitoring model, TensorRT-ready |
- DMS-MultiTask — lightweight driver monitoring model with 1.14M parameters and TensorRT-ready deployment
- VLM Data Closedloop — VLM-assisted hard-example mining, annotation, and augmentation pipeline for cabin monitoring
- Huazhou Dictionary — offline WeChat mini program for Huazhou dialect lookup with 4,310 entries
Python · TypeScript · C++ · FastAPI · LangGraph · PostgreSQL · Docker · PyTorch · TensorRT · MCP