Han HU (胡涵)
Hi there! I am Han HU (胡涵), a graduate student in Information Engineering at the Technical University of Munich (TUM) in Germany.
My research and engineering interests center around Large Language Models (LLMs), Agentic AI Systems, Distributed Training & High-Throughput Inference, and Embodied Multimodal Grounding. I am passionate about bridging theoretical machine learning models with high-concurrency, low-latency production engineering.
Currently, I serve as a Research Assistant at TUM focusing on LLM multi-turn complex instruction tuning (SFT), human preference alignment (DPO/RLHF), and distributed training across multi-node multi-GPU clusters (LRZ HPC). Previously, I was an Algorithm Intern at the Chinese Academy of Sciences (CAS) Collaborative Sensing Institute working on 3D-LLM spatial grounding.
Research & Technical Interests
LLM & Alignment
Multi-turn complex SFT, DPO/RLHF preference alignment, LLM-as-a-Judge automated evaluation, and anti-hallucination instruction design.
Agentic Systems
Autonomous task planning (ReAct, Plan-and-Solve), Tool Use / Function Calling, self-reflection loops, and sandboxed execution environments.
Distributed & High Concurrency
Multi-GPU training (DeepSpeed ZeRO-3, Slurm), and high-throughput low-latency inference serving with vLLM (PagedAttention, Chunked Prefill).
Embodied AI & 3D Grounding
Multimodal 3D-LLM, spatial instruction following, cross-attention geometric feature alignment (PointNet++, FPFH), and structured World Models.
News & Updates
- 2026.04 Joined Technical University of Munich (TUM) as a Research Assistant working on LLMs, SFT pipelines, and Slurm distributed training on the LRZ Supercomputing cluster.
- 2026.02 Completed development of the Self-Evolving Code Repair & Evaluation Agent based on Gemini, ReAct, and Docker sandbox isolation, achieving EV-Micro 8.50% and EV-Macro 23.03%.
- 2025.07 Completed the Real-Time Streaming Voice Agent project in collaboration with SAP, optimizing end-to-end full-duplex conversational latency from 3.4s down to 0.9s.
- 2025.04 Began Master of Science studies in Information Engineering at the Technical University of Munich (TUM), Germany.
- 2025.03 Concluded algorithm research internship at Suzhou Institute of Nano-Tech and Collaborative Sensing (CAS), focusing on Embodied AI and 3D-LLM spatial grounding.
- 2024.08 Graduated with First Class Honours B.S. in Data Science from University of Liverpool & Xi'an Jiaotong-Liverpool University.
Featured Projects
Real-Time Streaming Voice Agent in Interactive Scenarios (SAP Collaboration)
Engineered an end-to-end full-duplex conversational pipeline: streaming ASR → LLM intent reasoning → streaming TTS. Leveraged asyncio, WebSockets, Ring Buffers, and vLLM Chunked Prefill with INT8 quantization, achieving sub-second first-frame audio latency (<0.9s) with dynamic barge-in handling.
Self-Evolving Code Repair & Evaluation Agent based on LLMs
Built a Gemini-powered autonomous code repair agent with multi-stage closed loops (Intent Analysis → Dependency Localization → Patch Generation → Sandbox Execution → Error Reflection). Features containerized multi-agent execution with Docker, Celery orchestration, and automated Pytest feedback loops.
Let's Connect & Collaborate
Always open to exciting opportunities in LLMs, AI agents, and systems research.