I work on LLM post-training and RL. My working thesis: general-purpose models are commoditizing — the value is in adaptation. Every organization sits on proprietary data, structures, and constraints that off-the-shelf intelligence doesn't understand, and post-training is how you mold a model to fit. The real world's signal is rich and waiting to be captured.

Right now, I'm finishing research applying post-training techniques (reward design, GRPO, distributed training) to open-source models for materials science. The co-first-author paper is being written.

• SymbolicGym — symbolic reasoning environments for training RL agents, inspired by Gymnasium and formal methods
• Neural_Network_Sat_Guided_Design — using SAT solvers to generate neural network initial weights that satisfy constraints like orthogonality

I keep debugging notes from my training runs at jmadison.me — the weird failures are where the learning is:

• Why is my GRPO Loss 0?
• Why Is My Model Suddenly Speaking Thai?
• Why Does My GPU Utilization Keep Dipping?
• Speaking Thai, Part 2: Four Frozen Embeddings

Internships at AWS (Consultant), Cloudflare(Product), Datadog(Product), and Microsoft(Product + SWE), with an academic background in cybersecurity, privacy, and intelligence.

P.S. - Reward Hacking is just adversarial behavior with a KL penalty.