Machine Learning • Data Science • AI Engineering & Safety
I build AI systems from start to finish: multi-agent document pipelines, fine-tuned and calibrated transformers, LLM evaluation harnesses, and behavioral prediction models.
Madison, WI · Open to relocation linkedin.com/in/jackjburleson github.com/Exios66
Now
M.S. candidate in Psychology: Data Science in Human Behavior at UW–Madison (expected Dec 2026). My capstone, Digital-Mailroom, is built with industry partner American Family Insurance. It looks for ways to cut the cost of AI inference and improve how documents are read, sorted, and managed.
Featured Projects
Digital-Mailroom
Capstone · American Family Insurance
Five Python packages and one LangGraph pipeline that joins legal-document routing, prompt experiments, deterministic scoring, and deployment, in a reproducible Docker environment.
Mailroom-ML
Training, evaluation & calibration
A calibrated ModernBERT classifier with a guarded fast path and an LLM fallback. Includes ONNX CPU serving and a Modal fallback; evaluation coverage is not a production routing rate.
Neuron Atlas
Mechanistic interpretability · live browser model
Explore and steer all 8,192 neurons of TinyStories-8M in the browser, with random-direction controls and replayable findings. Larger Qwen studies are recorded offline.
Wave Laboratory
Side project · simulation & graphics
A real-time 3D ocean and ship-motion sandbox that runs in the browser, built on spectral wave physics.
What I’m interested in
I’m interested in AI systems that know when they are unsure. That means calibrated classifiers, evaluation you can trust, and pipelines that pass hard cases to a stronger model or a person instead of guessing. Outside of that, I like turning physics into interactive software, and I’m interested in how people think, which is what my psychology training is about.