PhD candidate at National Taipei University of Technology working on multi-agent deep reinforcement learning for UAV-assisted networks, reconfigurable intelligent surfaces (RIS), and space-air-ground integrated networks (SAGIN) — with a hands-on focus on GPU-accelerated PHY (CUDA, cuSolver/cuBLAS, NVIDIA Sionna). I also apply the same RL/ML toolkit to systematic trading: regime detection, risk control, and production ML. 🏆 4.00 / 4.00 GPA

Specialty domains: OFDM PHY · MIMO · LDPC · 5G NR · RIS · O-RAN · Federated Learning · MADDPG · PPO · CUDA kernels · Systematic trading

End-to-end OFDM PHY + AI-RAN stack where every claim is runnable. Real CUDA C++ MMSE channel estimation (custom kernel + cuSolver Cpotrf/Cpotrs + cuBLAS Cgemm), a Sionna 5G LDPC + TR38.901 TDL BLER link, and link adaptation (OLLA / model-based greedy / model-free learned policy) driven by the measured BLER curves. Result: 133× vs NumPy on an RTX 4090 (verified to ~1e-4 vs a CPU reference) · model-free policy matches a tuned OLLA / greedy from ACK/NACK feedback alone, at the lowest BLER · figures, green CI, and a NVIDIA_REVIEW.md design/verification guide

A CUDA kernel-optimization case study on the MMSE-apply complex GEMM: naive → shared-memory tiled → cuBLAS, benchmarked on an RTX 4090, with an Nsight Compute profiling methodology. Keeps a NumPy/CuPy reference and checks correctness against a CPU baseline. Result: clean naive→tiled→cuBLAS performance breakdown · GPU == CPU to ~1e-4 · CUDA compile-checked in CI · the GPU-kernel companion to the flagship

Standalone deep-RL study for 5G NR link adaptation: self-contained PyTorch PPO vs an OLLA industry baseline, 28-index MCS table (3GPP TS 38.214), non-stationary SNR with mobility/handover scenarios. Result: PPO learns a competitive policy from scratch (~3 min CPU training), evaluated fairly head-to-head vs OLLA · 15/15 unit tests

Federated learning for CSI feedback compression: CsiNet autoencoder + FedAvg under non-IID channel statistics, aligned with the 3GPP Release 18 AI-RAN study item. Result: FedAvg matches centralised (~−2 dB NMSE) and beats local-only by ~2 dB · 16/16 unit tests

• ris-beamforming-optimizer — RIS phase optimization (manifold + deep-learning methods)
• oran-resource-allocation-xapp — O-RAN xApp-style resource scheduling with DRL

An honest, reproducible benchmark for intraday order execution on real crypto LOB — classical (TWAP/VWAP/POV/Almgren–Chriss) vs deep RL (PPO/SAC) vs an LLM meta-controller, under one friction model (fees, queue, partial fills, adverse selection). Result (the honest null): no ML method beats a simple classical schedule on BTC or ETH — on ETH, deep RL is significantly worse (PPO, p = 0.035). The benchmark caught four simulator artifacts that had faked a ~2 bps edge; a training-length ablation rules out undertraining · paired-bootstrap significance · 13 unit tests · CI

Regime detection, risk allocation and live health monitoring — the open-sourced production layer of a systematic book. Six-indicator regime classifier (BULL/NEUTRAL/BEAR/CRISIS) with hard crisis gates, inverse-volatility weighting with portfolio vol targeting, and a monitoring battery with a trailing-drawdown kill-switch. Result: explainable-by-construction regime calls · fail-safe tradeable flag for automated halts · 9/9 unit tests · CI

Multi-asset systematic trading system on QuantConnect: cross-asset momentum, regime detection, volatility targeting and crisis routing, iterated through 11+ versions under strict out-of-sample validation with bias auditing (look-ahead, weight-cap, vol-estimation pitfalls). Status: active paper-trading track record · public repo contains a representative architecture sample

Systematic factor research on a professional simulation platform; first passing alpha cleared the platform's evaluation thresholds.

8+ IEEE publications in AI-native wireless networks:

• Hybrid federated learning with MADDPG for UAV-assisted access networks
• Reconfigurable intelligent surface optimization for 6G
• SAGIN architectures with LEO (Starlink) integration
• Channel estimation and beamforming for next-gen PHY

🔗 ORCID

• 🎓 PhD candidate, National Taipei University of Technology — 4.00 / 4.00 GPA
• 🟢 NVIDIA NGC 6G Developer Program — Member, 2026 cohort
• 👨‍🏫 Advisors: Prof. Hsin-Piao Lin · Assoc. Prof. Rong-Terng Juang

📧 Email · 💼 LinkedIn · 🎓 ORCID 📍 Taipei, Taiwan · 🇫🇷 🇬🇧 🇹🇼