Sean Chuang — AI Engineer Portfolio

English | 🌐 繁體中文

Hi, I'm Sean Chuang (莊家賓), an engineer focused on building and shipping AI applications. Below are projects I built end-to-end — model training, cloud deployment, and live demos. Feel free to click in and try them.

🧑‍💻 Tech Stack Overview

Category	Technologies
AI / Self-hosted LLM · VLM	Gemma-4-26B, Qwen3.6, Qwen3-VL, MedGemma, Llama-3.2-Vision, Qwen2.5-7B / 14B, llama.cpp / llama-server, vLLM,TensorRT-LLM, GGUF / 4-bit quantization, AWQ / GPTQ / FP8(ModelOpt)
Agent Architecture	Planner / Executor (ReAct) / Verifier, MCP (Model Context Protocol)
RAG / Retrieval	BM25 (tuned), bge-m3 (dense), RRF hybrid fusion, bge-reranker-v2-m3, NVIDIA NIM reranker
LLM Evaluation	LLM-as-judge (faithfulness), Cohen's κ calibration, stratified calibration sets, recall@k / hit@k / MRR, two-way human audit of the judge
Trustworthy AI / Uncertainty	Conformal Prediction (statistical guarantees), Selective Prediction (abstain & defer to humans), cross-model disagreement, Risk–Coverage / per-label AUC
Deep Learning Frameworks	PyTorch, TensorFlow, ONNX, HuggingFace transformers
Computer Vision	ViT, EfficientNetV2 + CBAM, YOLOv8 Nano (Ultralytics), OpenCV, TorchXRayVision
Software Quality / Testing	pytest, hypothesis (property-based), mutation testing, mypy strict, ruff, pip-audit, GitHub Actions CI
Training & Explainability	Focal Loss, Mixup, CutMix, Grad-CAM, Score-CAM, SHAP, Confusion Matrix
Healthcare Data Standards	HAPI FHIR R4, MedAgentBench, MIMIC-CXR / CheXpert
Backend / API	FastAPI, Flask, Nginx, Uvicorn, Gradio
Cloud / Deployment	AWS EC2 / RDS, GCP Compute Engine / Cloud Run, Docker
System Integration	LineBot, MySQL, HTML / JS
Tools / Languages	MobaXterm, Git, Python, WSL2

📂 Project 2: Choosing a RAG Configuration on a Single GPU (SelectRAG)

🏆 Key Results

👉 Answers "which RAG setup should we run?" with controlled experiments on a single RTX 4090: retrieval, inference engine and numeric format are changed one axis at a time. Hybrid + reranker lifts recall@3 from 0.808 to 0.941, and GPTQ-Int4 serves 2.36× the throughput of bf16. Even the LLM judge used for scoring was hand-audited in both directions, confirming quantization added no measurable errors.

🔗 Full technical details (GitHub Pages)

🔗 Source code (GitHub)

▶️ Watch the Web Demo

🧰 Technical Highlights

A decision, not a leaderboard: each axis gets a recommended pick, its cost, and a boring-but-safe alternative.
Controlled experiments: a 100-question human-labelled eval set frozen by hash; every number reruns with one command.
Strong baseline first: BM25 was tuned (recall@10 0.829 → 0.905) before comparing it with dense retrieval. Dense alone then added only 0.001; the real gain came from combining the two.
The engine is not the variable, the numeric format is: vLLM and TensorRT-LLM are within 10%, while W4A16 quantization gives ×2.36 throughput and 2.4× KV-cache headroom.
Auditing the instruments: the judge's κ fell from 0.845 on a random calibration set to 0.488 on a stratified one. A two-way hand audit showed its errors are symmetric across formats, and caught a fourfold dose error it had marked as supported.
Honest reporting: nine "expected to win, lost" results are all recorded, including two of three prompt revisions falsified by measurement.