AI infrastructure and AI agent engineer.
I build and study systems around LLM serving, retrieval-augmented generation, agent workflows, evaluation, and production-oriented AI tooling. My current focus is turning LLM capabilities into reliable software systems: serving backends, retrieval pipelines, tool-using agents, benchmarks, and operational scripts.
- LLM serving, inference benchmarking, and deployment workflows
- Retrieval-augmented generation systems for domain knowledge and code assistance
- AI agent workflows with tools, memory, planning, and task execution
- Evaluation, observability, and diagnostics for LLM applications
- Reproducible engineering: scripts, benchmarks, configs, and deployment notes
| Project | What it shows | Stack |
|---|---|---|
| vllm-ai_infra | Vehicle diagnostic knowledge retrieval and code-generation platform skeleton | Python, LLM, RAG |
| TwinForge | Learning platform for LLM training, inference, and local AI infrastructure | Python, CUDA, LLM infra |
| AI Agent Systems | Tool-using agent workflows for search, code assistance, automation, and task execution | Python, LLM, agents |
| Evaluation & Benchmarks | Scripts and experiments for measuring LLM serving behavior, latency, and reliability | Python, Shell, benchmarking |
I prefer small, measurable systems over demo-only code:
- benchmark before tuning
- keep prompts, tools, configs, and serving scripts versioned together
- document runtime assumptions and system dependencies
- make failure modes visible through logs, probes, evaluations, and repeatable tests
- design agents around clear tool contracts instead of opaque prompt-only behavior
PythonPyTorchvLLMRAGAI AgentsCUDADockerLinuxShellFastAPILangChainLlamaIndex
- GitHub: @heurry
- Interests: AI infrastructure, LLM applications, RAG, AI agents, evaluation, and automation