About me
I’m Yuchen Li, an AI engineer working on large language models, from training and fine-tuning to LLM applications in production.
In my current and previous roles I’ve built Graph RAG and multi-agent systems, including a self-hosted Qwen3 deployment on vLLM and a Text-to-SQL analytics PoC for accounting data.
- Built a Llama-style Transformer from scratch (GQA, RoPE, SwiGLU) and served it with FastAPI.
- Fine-tuned Qwen2.5-7B with LoRA on 790K Chinese medical Q&A pairs on a single 24GB GPU.
- Built a RAG chat backend with hybrid search and streaming responses.
Stack: PyTorch · Transformers · LoRA / Unsloth · vLLM · FastAPI · LangChain · Neo4j · Elasticsearch
Education: M.Eng. in Artificial Intelligence (2025) · B.Eng. in Civil Engineering (2022)
Contact
- Email: manshuer.li@gmail.com
- GitHub: bigrh0917
Category
Manshu LLM: A Character-Level Transformer Language Model Built from Scratch
A Llama-style decoder-only Transformer (GQA, RoPE, sliding-window attention, SwiGLU) built from scratch in PyTorch, served via FastAPI.
Manshu RAG: Knowledge-Augmented Chat System
A FastAPI RAG backend with async document ingestion, hybrid vector + keyword search in Elasticsearch, and streaming LLM chat over WebSocket.
Manshu Finetune: Chinese Medical QA LLM Fine-Tuning with Unsloth
LoRA fine-tuning of Qwen2.5-7B-Instruct on 790K real Chinese medical Q&A pairs with Unsloth, trained end to end on a single 24GB GPU.