AI Engineer · Agent Developer · Open Source Builder
AI Engineer at Huawei, owning quantization and adaptation of multi-modal diffusion models and LLMs on Ascend chips. I build at the intersection of RAG, ReAct agents, and low-level inference optimization — turning ideas into verifiable, reusable engineering pipelines.
- 🤖 AI Engineer based in Shanghai, China — specializing in LLM inference, Agent systems, and model quantization.
- 🐱 Cat lover · ⚽ Football · 🎸 Guitar & 🎹 Keyboard · 📷 Photography.
✈️ Travel · ♟️ Chess · ⌨️ DOTA 2 — always up for a good game or adventure.
- 🦁 Columbia University — M.S. in Data Science (Sep 2023 - Dec 2024)
- Focus: Generative AI & LLM Applications. Coursework: Computer Vision, NLP, Statistical Modeling, ML in Practice.
- 🔱 University of California San Diego — B.S. in Data Science, Minor in Economics (Sep 2019 - Mar 2023)
- GPA 3.9/4.0 · Cum Laude · Provost Honors (2019-2023)
- 🏫 The Derryfield School — High School Diploma (Sep 2017 - Jun 2019)
- 📚 Linfield Christian School — High School (Aug 2015 - Jun 2017)
-
🤖 AI Engineer at Huawei Technologies Co., Ltd. (Aug 2025 - Present | Shanghai)
- Multi-modal Diffusion Model Quantization & Optimization: Led quantization adaptation for Flux, Wan 2.2, and Hunyuan models on next-gen Ascend chips. Achieved 1.5× inference speedup for Wan 2.2 & Hunyuan, and 2×+ for Flux (exceeding the 1.6× target) through operator-level profiling, precision tuning, and fused operator integration. Built automated DiT-block profiling analysis tool (pandas-based), saving ~50% analysis time and adopted as team standard across 5+ projects.
- Qwen3.5 End-to-End Adaptation (Model Owner): Fully owned Qwen3.5 adaptation on vLLM framework for next-gen Ascend chips — covering functionality validation, quantization, operator fusion, PD separation deployment, and end-to-end performance tuning. Coordinated cross-team efforts across operator, framework, and quantization teams to ensure on-time delivery.
- SGLang MXFP Quantization: Contributed MXFP4 & MXFP8 quantization support to SGLang across three model scenarios — Diffusion, Dense LLM, and MoE LLM. Leveraged Claude Code for AI-assisted development, completing the originally estimated 2-month workload in ~1 week (~75% efficiency gain), validating AI-assisted engineering for complex framework-level tasks.
- Engineering Efficiency & Team Impact: Early adopter and advocate of AI-assisted development (Claude Code) within the team — led 2 internal tech talks, authored the team's AI-assisted coding guide. Built a web-based server whitelist management system to improve team resource utilization. Contributed technical documentation and led multiple internal knowledge-sharing sessions on quantization, performance tuning, and AI-assisted development.
-
🧬 Data Analyst at Columbia University Medical Center (Nov 2023 - Dec 2024 | New York)
- Designed end-to-end neuronal activity analysis pipeline for in vivo recording data. Quantified temporal relationships between neural activity and feeding behavior through statistical modeling, supporting identification of key neuropeptidergic neurons regulating satiation.
- Co-author on "Brainstem neuropeptidergic neurons link a neurohumoral axis to satiation" (under review at Cell). Built automated data processing scripts adopted for subsequent similar studies.
-
🎉 Head of Development at UCSD CSSA (Mar 2020 - Mar 2023 | San Diego)
- Built the organization's official website with Vue.js, Vite, and Flask. Created programming tutorials for 15 team members and designed an internal knowledge-sharing portal.
- UCSD CSSA Website · Department Portal
| Level | Skills |
|---|---|
| Expert | Python (Pandas / NumPy / Matplotlib), PyTorch, LLM Inference & Serving (vLLM, SGLang), RAG (LangChain, Pinecone), Agent Development (ReAct, Function Calling, MCP) |
| Proficient | Model Quantization (MXFP / INT8), Prompt Engineering, OpenAI / DeepSeek API, Streamlit, OpenCV, Scikit-learn, SQL, Flask, asyncio / httpx, GitHub REST API |
| Familiar | TensorFlow, Vue.js, TypeScript, D3.js, Java / Spring, Redis, MongoDB, Docker, Git, Linux |
| Project | Description |
|---|---|
| react-agent | Production-grade ReAct Agent built from scratch — zero framework dependencies, DeepSeek API driven. Five-layer architecture with per-tool circuit breakers, exponential backoff with jitter, and graceful degradation. Chain-of-thought visualization via DeepSeek reasoning_content. |
| TD LLM Streamlit | Gen AI Executive Advisor for TD Bank — RAG-powered strategic Q&A with GPT-4o + LangChain + Pinecone Serverless. Semantic sliding window chunking, multi-stage retrieval, hallucination guard, and truLens LLM-as-Judge evaluation. |
| SGLang Quant Eval | MXFP8/MXFP4 quantization on SGLang for Huawei Ascend NPU — covering Diffusion, Dense LLM, and MoE scenarios with both online quantization and offline pre-quantized weight inference. |
| DOTA 2 Drafting Backend | ML backend for DOTA 2 hero recommendation — 30K+ matches, 821-dim feature engineering, PyTorch / XGBoost / Random Forest ensemble. 51.8% overall accuracy, 58.3% on known counter subsets. |
| DOTA 2 Drafting Frontend | Vue 3 + Element Plus frontend with captain-mode hero selection interface for DOTA 2 match prediction. |
| Personal Website | Modern responsive portfolio built with Vue 3, TypeScript, and Element Plus — VS Code-inspired design with i18n and dark/light themes. |
| agent-building | Progressive 4-week AI Agent system build — ReAct loops, MCP protocol tooling, resilient fault tolerance, and LangGraph state machines. |
| Iris Recognition | Deep learning-based iris recognition with PyTorch. |
| Climate Analysis | 100 years of climate data analysis and global warming visualization with Python & D3.js. |
- Chowdhury, S., Kamatkar, N. G., Wang, W. X., Akerele, C. A., Huang, J., Wu, J., Kane, C. M., Bhave, V. M., Huang, H., Wang, X., & Nectow, A. R. (under review). Brainstem neuropeptidergic neurons link a neurohumoral axis to satiation. Cell.
- Wu, J. (2022). A WeChat Application Combined with Convolutional Neural Network for Skin Cancer Recognition. International Conference on Artificial Intelligence in Education (ICAIE).
- 🌐 Website: tallmessiwu.com
- 📧 Email: junlin-wu@foxmail.com
- 🐙 GitHub: TallMessiWu
- 📸 Instagram: @tallmessiwu
- 📝 Weibo: @tallmessiwu