Building infrastructure, tooling, and AI systems to remove friction from technical workflows.
Interested in Research Engineer, ML Engineer, and MTS roles.
Jordan Madison
Hola! I play Rocket League and spend more time than I should wondering how many bits it takes to reach the optimum of a loss curve.
For the past year and a half, I have been building post-training infrastructure and evaluation systems for open-source LLMs applied to autonomous materials discovery. I started off assuming building these systems was tough and I quickly ran into the reality the work is hard and grueling.
The 3 am prayers to the HPC gods to get my jobs off the queue, deciphering why my model was speaking Thai, and wasting hundreds of H100 GPU hours on a silent chemistry tool call error I didn't realize because I was too busy trying to graduate are a tiny subset of a humbling, yet eye-opening experience I'll never forget. Every day tested my sanity while performing research. I was finishing up the last year of my degree, but those late nights were worth it when Qwen3-0.6B output the hypothesis and modification tags, the environment property evaluation tool worked, and I witnessed my first successful modification. I am not a material scientist by trade, I am a computer scientist, so I was just elated the entire system finally worked.
Ever since, I have iterated on improving the infrastructure to support my research and develop a deeper intuition about LLMs and RL at scale. From generating custom training datasets for SFT to fitting inference, training, and scientific calculators on a single H100 to trying to make sense of multi-step trajectories to gauge model behavior, I fell in love with bringing all of these concepts together and thinking about how it all plays out at scale not only in materials science, but in all domains.
I'm eternally grateful for the research opportunity I've had this past year. It put me in a position to take a risk on an area I knew little about, exposed me to a ton of amazing lines of thought, and opened my eyes. There is a lot more work that needs to be done to build systems that help people iterate faster, explore expensive ideas, and push the boundaries of discovery across all domains. That is the kind of work I want to do.
Selected work
-
Post-training research
Co-first-author paper in preparation on physics-certificate GRPO for band-gap engineering. Across four open-weight language models, higher rewards did not consistently improve exact task success.
View research → -
Cloudflare Radar
Led 0-to-1 product work that turned global network telemetry into a public internet intelligence platform shipped in 12 weeks.
View experience → -
GRPO debugging series
Field notes from a year of RL post-training: zero loss, tokenizer failures, frozen embeddings, and GPU utilization sawtooths.
Read the writing →
Recent writing
-
An application of self-improving harnesses to limit order placement.
#design , #finance , #investment , #agent , #meta
-
A breakdown of an RL environment I designed after misreading a take-home assignment. The design is built around my reverse-enginering background and I liked it enough to write it up here.
-
A LoRA debugging story about four embeddings that never got to learn anything. If you add tokens and freeze the base model, their embedding and LM-head rows still need to train.
#post-training , #sft , #lora , #tokenizers , #ai-assisted
-
Reading GPU telemetry on a colocated GRPO run: periodic utilization dips, memory-access spikes, and a bottleneck story of mine that didn't survive contact with the code.
#rl , #post-training , #grpo , #vllm , #ai-assisted
-
A debugging story about special tokens, GRPO, and the skip_special_tokens flags scattered across your RL stack. When your SFT'd model nails the format in eval, then falls apart the moment it enters the GRPO rollout.
#rl , #post-training , #grpo , #tokenizers , #ai-assisted