Software engineer. About nine years on the backend of an institutional money-market trading platform in regulated fintech. Before that, US Army satellite communications operator, and before that, corporate security operations. Bootcamp-trained, not degree-trained. Denver, Colorado.
On the side I build LLM systems and run small research programs the way I wish more software were built: pre-registered hypotheses, kill criteria written before the run, and negative results kept in the log.
All of the projects below were built with AI coding assistants in the loop. The questions asked, the evaluation design, the calls about what to trust, and the decisions to kill things are mine. Ask me about any of them.
| Project | What it is | Why it might interest you |
|---|---|---|
| teatime | A map of the scientific literature that an LLM navigates to find topically distant, structurally close papers. Five layers from LLM-extracted "mechanism cards" to an MCP server. | A date-capped pre-discovery benchmark: 23 historical cross-domain discoveries, each with a corpus built before the discovery. Mechanism search reaches hit@100 of .51 to .59 against .275 for vanilla similarity, with four named leakage channels and planted-bridge controls. The findings log corrects itself in public. |
| millennium | Computer-assisted blow-up numerics for 2D Boussinesq, run as a pre-registered program on rented H100s, with an adversarial review folder. | 280 commits of disciplined numerical work with per-campaign verifiers that must pass before a result is accepted. Nothing here claims a proof. |
| VNR | A lab notebook on the ARC-AGI-1 to ARC-AGI-2 compositional-generalization collapse. | Fourteen logged experiments. Most outcomes are KILL or INCONCLUSIVE, recorded honestly. |
| cascade | Local-first LLM router: three local tiers on one GPU, three cloud tiers above, and an answer-then-verify judge that escalates. | Grew out of a simulation study on whether volunteer compute could replace centralized serving. The coding agent built on it edits its own source. |
| aether-engine | A Rust game engine where worlds and mechanics are data packages, patched live over SSE from a browser Studio. | Nine crates, about 10k lines, unit-tested. |
| env-lab | An LLM directs three factories (assets, density, lighting) to build game environments in Unreal and Blender, judged by a reference-anchored critic with human-owned ratings. | The model never makes art. It directs, and a human rates. |
| greenrest | Godot RPG prototype with classless progression: every class is a track any character can level. | Playable vertical slice, CC0 art, automated screenshot harness. |
| RecipeBox | Photograph a recipe card, a vision model transcribes it, it lands in your group's shared cookbook. Next.js, Supabase with row-level security. | Started as digitizing 316 of my grandmother's cards. |
| undesigned-coincidences-paley-blunt | Public-domain restorations and modern-English translations of Paley's Horae Paulinae and Blunt's Undesigned Coincidences. | OCR-verified, freshly typeset, released CC0. |