Haofei Sun

Engineer working across LLM systems and embedded ML, graduating M.S.(Quit Phd) Computer Science at UT Arlington in December 2026. I build systems at the boundary between real-time firmware and modern ML — from Zephyr RTOS drivers on nRF54L15 to physics-informed deep-learning models reconstructing sub-Nyquist RF signals.Open-source (2026): merged fixes into leading LLM-infrastructure projects — SGLang (~29k★ serving framework; multi-tenant batching crash + co-authored a metrics fix), LiteLLM (50k★ gateway; diagnosed a multi-region Vertex AI routing bug and contributed the merged regression suite), and LangChain (repo-wide prompt-encoding fix). I also maintain RepoAgentBench, an open-source coding-agent benchmark on PyPI.Current research (UTA, Aug 2025 – present): built real-time Zephyr firmware acquiring BLE RSSI at 77 kHz (3× prior published results) with <0.01% drop rate, and trained a physics-informed PyTorch model on NVIDIA B200 GPUs that reconstructs complex-valued RF signals from aliased scalar measurements at 0.986 R² on chirp recovery. The firmware and the model are one pipeline — that's what I care about.Prior work: two years of mixed-signal / ADC design at Xidian (Verilog + MATLAB/Simulink, 2 Chinese patents accepted). Python ETL pipeline at VisualComm. Sim-to-real RL on Franka Panda and xArm in Isaac Lab since late 2024. Deployed LLM agents on Streamlit + HuggingFace (SmartStudy, adaptive OODA-loop tutoring) and a full-stack LangChain + MCP agent (Agentic Weather).Strengths I'd bring to a team:- Open-source contributor to production LLM infrastructure (4 merged PRs)- End-to-end: RTL → firmware → DMA data path → Python preprocessing → PyTorch training → deployed inference- Typed, testable agent code — not just prompt-hacking- Comfort with low-level C and Verilog when the problem actually needs it- Track record of turning research prototypes into things that runOpen to full-time SWE, ML Engineer, Embedded Software, and AI Engineering roles starting early 2027 (OPT eligible, STEM extension). Happy to connect with other Mavs, embedded + ML engineers, and anyone building at the hardware/ML boundary.haofei.sun@uta.edu · github.com/HumphreySun98