Projects
Built nights and weekends, outside the day job. Each one exists because I wanted the answer.
01
"You upgraded your model. What broke?"
Statistically rigorous drift detection between LLM endpoints. It scores eval suites with paired significance tests, prices every verdict in cost per correct answer, and runs a scheduled observatory that watches live endpoints for silent model swaps.
I use it to document drift between frontier releases from the outside. One finding: a silent tokenizer change raised a model's cost per correct answer 35% while accuracy barely moved.
02
Consumer scam and fraud detection using multimodal LLMs. Paste the suspicious thing, get a read on it.
Designed, built, and shipped solo.
03
"Your AI agent needs a warrant."
The permission layer and the paper trail for AI agents taking real-world actions. You write allow, ask, and deny rules in a Markdown file the agent checks before it acts. Every action lands in a hash-chained audit record that breaks if edited.
Early and in the open, version 0.2. The CLI, MCP server, and warrant format work today. The warrant card on this site is docket's idea, pointed at you. Companion essay: Your AI agent needs a warrant.
04
A framework and diagnostic for measuring enterprise GenAI ROI.
From my Georgetown capstone, built on a 44-case empirical dataset.
05
Frees a consumer AI voice recorder from its vendor cloud. I reverse-engineered the BLE protocol, so recordings transfer straight to your machine and transcription, speaker ID, and summaries all run locally.
The vendor did not publish the protocol, so I read it off the wire.
Also on GitHub