Shawn Wei

AI and backend engineer with 6+ years focused on LLM inference, fine-tuning, and multi-agent systems. Most recently I architect a production LLM inference stack: Triton + TensorRT INT8 on A800 80G at 256K context, which doubled throughput and cut GPU cost 35% with zero new hardware, plus vLLM serving with INT8 quantization and MTP speculative decoding. I fine-tuned Qwen3.5-9B for government documents (SFT + GRPO, 10K+ distilled samples). On the agent side I built a production VLM auto-clip pipeline (GLM-4.6V + Qwen3-VL metadata extraction, MinIO + OpenSearch vector search, SHA256 dedup, Docker deployment) and a DeepSeek V4 multi-agent system for enterprise automation (planner/executor/reviewer patterns). I also independently build and run two live products - fathomsearch.xyz (a multi-engine search API) and nsfwgram.cc (a full-stack Next.js platform) - which means reversing non-standard APIs, Playwright/CDP browser automation, Cloudflare/PerimeterX fingerprint bypass, and keeping messy external behavior stable and monitored. Stack: Python (primary), TypeScript/Node.js/Next.js, PostgreSQL, Docker, Caddy, BunnyCDN, Tailscale; vLLM, Triton, TensorRT, SFT/RLHF/GRPO, LoRA, RAG, MCP. 20+ models on HuggingFace (shawnw3i), 25K+ downloads. C1 English - async docs, Slack, and code review all in English. Kunming, China (UTC+8); 2-3h overlap with US-Eastern evenings or EU mornings. Looking for a senior remote AI/LLM inference or agent role where I can own a production serving or agent pipeline end-to-end.