Han HU (胡涵)

Han HU (胡涵)

M.S. in Information Engineering @ TUM | LLMs & Embodied AI

Curriculum Vitae

Han HU (胡涵)
Han HU 胡涵
Spring 2026
M.S. in Information Engineering · Technical University of Munich (TUM)
huhan606@gmail.com Munich, Germany github.com/huhan606 LinkedIn
Education Experience Projects Skills
Education
Technical University of Munich (TUM) 2025.04 – 2027.04
M.S. in Information Engineering German GPA: 1.6
University of Liverpool & Xi'an Jiaotong-Liverpool University (XJTLU) 2020.09 – 2024.08
B.S. in Data Science (Dual Degree) First Class Honours
Professional & Research Experience
Technical University of Munich (TUM) 2026.04 – Present
Research Assistant (LLMs & Agents) · LRZ Supercomputing Center
LLaMA-Factory DeepSpeed ZeRO-3 vLLM Slurm PyTorch
  • Instruction Tuning (SFT) & Domain Alignment: Designed a multi-turn complex SFT pipeline for vertical dialogue and task planning; engineered prompt templates and preference datasets for open-source foundation models (Qwen/Llama) using LLaMA-Factory, mitigating hallucination and boosting benchmark accuracy by +23%.
  • Preference Alignment (DPO/RLHF): Implemented preference alignment with LLM-as-a-Judge multidimensional scoring and fine-grained ranking, significantly enhancing communication quality, safety compliance, and multi-turn logical consistency.
  • Distributed Training (Slurm & ZeRO-3): Deployed multi-node multi-GPU clusters (A100/H100) via Slurm on LRZ HPC; integrated DeepSpeed ZeRO-3, FlashAttention-2, and gradient accumulation optimization, reducing per-GPU memory usage from 78GB to 52GB (-33%) and accelerating end-to-end throughput by +28%.
  • High-Concurrency Serving (vLLM): Built low-latency, high-throughput model serving pipelines with vLLM; leveraged PagedAttention, Continuous Batching, and Prefix Caching to optimize KV cache reuse, reducing long-context Time-To-First-Token (TTFT) by -45%.
Suzhou Institute of Nano-Tech, Chinese Academy of Sciences (CAS) 2024.09 – 2025.03
Algorithm Intern (Embodied AI & Spatial Grounding) · Collaborative Sensing Lab
Embodied Agent 3D-LLM PointNet++ PyTorch Spatial Grounding
  • Embodied Agent & Spatial Grounding: Researched 3D-LLM multimodal feature alignment using Cross-Attention and geometric embeddings (FPFH/PointNet++), resolving physical interaction ambiguities in complex spatial instruction following.
  • Multimodal Geometric Fusion & World Model: Combined RANSAC robust estimation, Gauss-Newton optimization, and ICP/NDT registration to counter occlusion and point cloud noise, cutting pose estimation error by -65% (RMSE 5.2mm → 1.8mm) and single-frame latency by -62.5% (120ms → 45ms) for upstream agent planning.
Project Experience
Real-Time Streaming Voice Agent in Interactive Scenarios 2025.05 – 2025.07
Lead Architecture · In collaboration with SAP
FastAPI WebSocket vLLM ASR / TTS asyncio
  • Full-Duplex Conversational Loop: Engineered an end-to-end streaming conversational pipeline (Streaming ASR → LLM Intent Reasoning → Streaming TTS) for natural, fluid anthropomorphic customer service dialogue.
  • Context Routing & Barge-in Handling: Developed sliding-window dynamic context stitching and intent routing modules, robustly maintaining long-horizon context consistency with seamless user interruption (barge-in) flow control.
  • Sub-Second Latency Optimization: Re-architected chunked transport with asyncio, WebSockets, Ring Buffers, and vLLM Chunked Prefill with INT8 quantization, slashing first-frame voice response latency from 3.4s to <0.9s.
View Project Case Study →
Self-Evolving Code Repair & Evaluation Agent based on LLMs 2025.10 – 2026.02
Autonomous Agent Research & Engineering
ReAct Docker Sandbox Django Celery Redis
  • Agentic Workflow & Tool Use: Built an autonomous code repair agent on Gemini featuring closed-loop localization, patch planning, sandboxed execution, and self-reflection (100% tool invocation success, 11.40% end-to-end task success in complex long-context environments).
  • Benchmark Evaluation & SOTA Performance: Evaluated against open-source SOTA baselines (Agentless, OpenHands) on automated benchmarks, achieving execution verification metrics of EV-Micro 8.50% and EV-Macro 23.03%.
  • High-Availability Sandboxed System: Built an asynchronous orchestration hub with Django REST Framework, Celery, Redis, and containerized Docker sandboxes with automated Pytest feedback loops, averaging only ~0.35M tokens per task.
View Project Case Study →
Technical Skills
Core Programming & Developer Tools
Python (Advanced) C++ SQL Linux / Bash Git Slurm (HPC) Docker

Proficient in Python for machine learning workflows and systems programming; experienced with Linux HPC clusters, shell scripting, containerized environments, and Git-based collaborative development.

LLM Algorithms & Alignment
Transformer SFT (LoRA/QLoRA) DPO / RLHF LLaMA-Factory Prompt Engineering (CoT) LLM-as-a-Judge RAG

Deep understanding of Transformer architectures; hands-on fine-tuning and alignment of open-source LLMs (Qwen/Llama); automated multidimensional quality evaluation and hallucination mitigation.

Agentic Systems & Multimodal AI
ReAct & Plan-and-Solve Tool Use / Function Calling Docker Sandbox Isolation Full-Duplex Voice Agents 3D-LLM Spatial Grounding PointNet++

Proficient in designing autonomous agent closed loops with self-reflection; full-duplex streaming conversational pipelines; multimodal spatial instruction following and 3D geometric grounding.

Distributed Systems & Inference
PyTorch DeepSpeed ZeRO-3 FlashAttention-2 vLLM (PagedAttention) FastAPI Celery & Redis WebSockets asyncio

Multi-node multi-GPU cluster distributed training; high-throughput low-latency inference serving; asynchronous architecture design with containerized sandbox isolation.