Portfolio →  ·  Email  ·  LinkedIn  ·  Medium  ·  繁體中文

From financial fraud detection and traffic-flow forecasting to AI explainability and fairness — I believe good models should not only be accurate, but also able to explain themselves.

Technology Consulting Consultant · AI/ML @ Ernst & Young Taiwan  ·  2025/07 — Present

Consultant in EY's Technology Consulting practice, with AI/ML as my hands-on specialty.

• 🤖   RAG Knowledge System — Leading architecture & development of an internal RAG-based knowledge management system (PoC), running locally on Llama 3.1 via Ollama.
• 🏥   Patient Health-Education RAG (in dev) — Building a RAG system with a healthcare institution that generates a personalized health-education sheet for each patient.
• 🛒   Retail ML Recommender — Real-time product recommendation engine for a major retail chain (PoC), combining user behavior with product nutrition signals, refreshed in 5-minute batches.
• 🏦   Banking Credit Scoring — Owning credit-scoring model development for a banking client. Completed SAS Viya certification and supported the client's platform onboarding.
• 🧪   AI Model Governance — Owning fairness, explainability and AI capability evaluation for enterprise clients (SHAP analysis + multi-dimension fairness audit).
• 🎓   EY Corporate Mentor × NTU Accounting — Mentored a 6–7 student team through a 16-week, 3-hour-per-week RAG project (Spring 2026 — completed, final reports delivered).
• 🎤   AI Ambassador Program Lecturer — Invited lecturer for the Digital Development Ministry's AI Talent & Rural Digital Enhancement program (AI Ambassador), via EY.

Master's thesis · Tamkang Data Science · Graduated July 2026

Training AI to detect misleading charts through red–blue adversarial self-play: Red generates attacks, Blue learns to catch them, both co-evolve via Dual-GRPO. The core result is mechanistic — the bottleneck is the model's decoding interface, not its representational capacity. On the same vision-language features, swapping generative JSON decoding for a sigmoid classifier head lifts balanced accuracy 0.57 → 0.70; fusing with a rule-based linter reaches 0.83. On synthetic adversarial charts, the 3.7M-parameter detector hits BalAcc 0.943 / AUC 0.997 — well ahead of general-purpose large models.

The out-of-distribution ceiling is characterized honestly rather than hidden: all 12 detectors plateau on wild charts, traced substantially to renderer shift.

_{↑ Play as the Red attacker: pick misleading techniques and see if the trained Blue Critic catches you.}

Hackathon finalist · DIGITIMES × AWS 2026

AI-driven crypto fraud detection across 63,770 users · 129,545 graph edges. HeteroSAGE + GAT heterogeneous GNN with stacking ensemble (XGBoost + LightGBM + CatBoost) on a 78-dim feature space. SHAP explainability and 4-dimension fairness audit baked in.

2nd Place (1st vacant), National Highway Comp. 2025 · Invited talk at Transportation Annual Conference

Dual-engine architecture combining MT-STNet (spatio-temporal GNN) with LWR shockwave theory for accurate congestion prediction and real-time alerts. RAG decision support adds explainability on top of prediction.

_{Invited talk at the 2025 Chinese Institute of Transportation Annual Conference.}

Best Popularity + Merit Award · Academia Sinica 2024

First application of Generalized Association Plots (GAP) to cyberbullying research. Combined with PCA, factor analysis and CCA, identified 5 user clusters and 4 behavioral factors from 672 respondents.

AI / ML  ·  PyTorch · PyG · GNN (HeteroSAGE / GAT) · RAG · LLM · Ollama · XGBoost · LightGBM · CatBoost · SHAP

Languages  ·  Python · TypeScript · Java · R · SAS · SQL

Frontend  ·  Vue 3 · React 18 · Tailwind · D3.js · Three.js

Backend & Infra  ·  FastAPI · Node.js · Spring Boot · Docker · AWS · Linux · MySQL

Statistics  ·  Multivariate Analysis · GAP · PCA · CCA · Factor Analysis

Specialized  ·  SAS Viya · MT-STNet · Fairness Audit

Certified  ·  Azure AZ-900 · AI-900 · DP-900 · SAS Base · ESG Junior Manager

• 📚   Claude Code Field Manual — a beginner-friendly zh-TW tutorial series (G1~G20, ongoing)  ·  iThome · 2026
• 📰   When 97% of Your Data Lies — Extreme Imbalanced Classification in Crypto AML  ·  Medium · 2025
• 📰   Building a Portfolio in One Day with AI — Claude Code Collab Reality Check  ·  Medium · 2025

_{Open to AI/ML collaborations, talks, and tech discussions.
chwei9181@gmail.com}