Selected Work

Post-training, evaluation, production ML, and RL research

LLM post-training and evaluation (Postman)

As a Senior AI Software Engineer at Postman and the DRI for model training and evaluation, I own the full loop of training a model and measuring it.

  • Post-training pipeline. Built a multi-phase fine-tuning pipeline on large-scale GPU infrastructure covering SFT, DPO, RLVR with GRPO, and reinforcement fine-tuning (RFT), with data-quality filtering and benchmark-driven checkpoint evaluation. Diagnosed and fixed a training collapse caused by catastrophic forgetting, and added safeguards against recurrence.
  • Evaluation benchmark. Designed an industry-grade evaluation benchmark and environment for frontier LLMs, covering multiple evaluation axes with deterministic validators, bootstrap-CI scoring, anti-contamination controls, and multi-trial averaging. When frontier models saturated the initial tasks, I rebuilt a harder suite that restored a wide discriminative spread for meaningful cross-model comparison.
  • Capability vs refusal. Found that safety-training-driven abstention was systematically understating true model capability, and introduced an abstention-adjusted metric that separates capability from refusal.
  • Agentic systems. Designed and deployed a next-generation agentic AI system and dynamic LLM routing strategies for a production enterprise platform.
  • Partnerships. Primary technical point of contact for AI partnerships and external model-evaluation collaborations.

Production ML (Bobyard)

  • Built and shipped a production VQA and RAG system (FastAPI, PaddleOCR, LightRAG, AWS) over a two-layer knowledge graph with per-project data isolation, serving semantic and relationship search across customer drawing sets.
  • Architected a representation-learning and embedding pipeline for 100-megapixel and larger engineering drawings (custom compression and detection models) and the distributed GPU/CPU infrastructure behind it, handling multi-gigabyte batches.
  • Earlier, as a computer vision intern, built zero-shot symbol detection and a Segment Anything Model (SAM) based method for detecting long thin lines in construction drawings.

Reinforcement learning research (Virginia Tech, PhD)

  • Decision-theory-guided deep RL for fast learning. Built a framework combining decision theory with PPO to mitigate the cold-start problem, achieving up to 184% higher initial reward and 53% more accumulated reward than standard deep RL. Code: github.com/Wan-ZL/DT-DRL.
  • Honey drones for drone surveillance. Modeled attacker and defender as A3C deep RL agents competing over signal strength, with parallel training for faster convergence. Code: github.com/Wan-ZL/gym-drones.
  • Defensive deception against advanced persistent threats. Used hypergame theory and deep RL for defensive deception against APTs (Foureye). The latest work in this line, on cyber deception for mission surveillance via hypergame-theoretic deep RL, was accepted to IEEE TDSC in 2026. Code: github.com/Wan-ZL/ARO-Foureye.
  • Uncertainty-aware human-AI teaming. Built a Vision Transformer pipeline with Evidential Deep Learning so the model can express uncertainty, across four U.S. Army-funded projects.