Applied ML & reinforcement learning · NYCU, Taiwan
Consultant on the AI & Data team within Technology & Transformation at Deloitte, applying machine learning and data science to business problems. Research background in reinforcement learning, sequence models, and multimodal ML from National Yang Ming Chiao Tung University (NYCU) — drawn to problems where a learned policy or representation has to hold up under distribution shift (financial markets, speech, biomedical signals).
🏛️ MacroHFT × Dynamic Hybrid — Adaptive multi-agent RL for crypto HFT
Extended the KDD 2024 MacroHFT memory-augmented hierarchical RL framework with a Dynamic Hybrid coordination mechanism that adaptively switches between soft-consensus and hard-expert-selection at the hyper-agent level, gated by a learned market-regime signal.
On minute-level ETHUSDT (held-out test set):
| Metric | MacroHFT baseline | Dynamic Hybrid (mine) |
|---|---|---|
| Total Return | 6.05% | 38.66% |
| Sharpe | 0.87 | 3.64 |
| Profit Factor | 1.42 | 4.08 |
Also explored Rainbow DQN and QR-DQN sub-agents as documented negative results — kept in the repo because the regime they failed in informed the Dynamic Hybrid design.
🏭 AMR Fleet Sizing & Scheduling — How many robots does a fab need?
Frames a semiconductor fab's material-handling bottleneck as an optimization problem: find the smallest AMR fleet (lowest capital cost) whose schedule still clears a month of wafer transport. A genetic algorithm does the scheduling, cross-checked against an exact Gurobi MILP baseline; an automated search returns the business answer — 6 AMRs (TWD 36M) on the demo instance.
The repo's signature move is intellectual honesty: a random-baseline check shows the GA beats naive assignment by only ~1–2% here — and explains why (no time windows, so load-balancing is trivially easy), turning a "my method won" demo into a clear-eyed read of when a metaheuristic is actually the right tool. NYCU GA course project; tests + CI.
🧬 Voice-Face Aging Prediction — Multimodal healthy-aging signal
Multimodal model combining voice and facial-expression features to predict markers of healthy aging.
🗣️ FSR-Challenge-2025 — Hakka ASR
Competition entry for the Formosa Speech Recognition Challenge 2025 (Taiwanese Hakka). Both tracks: Hanzi (CER) and Pinyin (SER).
Deep models for 12-lead ECG arrhythmia classification on the CPSC 2018 dataset.
🎧 shadowing-player — Listening practice tool
YouTube sentence-by-sentence blind-listening trainer for language learners — listen first, reveal subtitles after.
Working areas: deep reinforcement learning (DQN family, policy gradient, hierarchical RL), transformers & speech models, multimodal learning, LLM fine-tuning.
📧 hsiehk0214@gmail.com · 💼 LinkedIn