iridis.
2023– · Results committed
Tested whether image-derived skin color tracks Fitzpatrick type and Monk Skin Tone on open dermatology images, and found much of the image error is the camera, not the scale. A color matrix per device and dermoscope mode, fitted on held-out patients, recovers a colorimeter at ρ 0.915 and cuts the capture share of variance from 40% to 5.8%.
Read the write-up →lambent.
2023– · Results committed
Skin radiance metric from multi-region image features, developed independently on public data. Re-validated on 1,816 public dermatology images and rebuilt into a capture-robust, tone-neutral score (tone dependence ρ² 0.310 to 0.002).
Read the write-up →veridian.
2025– · Results committed
Literature review accelerator for any field: from a one-line topic and an inclusion statement it drafts searches, snowballs citations, screens every abstract with an LLM and maps the result. On human hair fiber science it screened 3,205 candidates to 1,199 in 5.7 minutes, with all 20 sampled inclusions in scope and 20 of 24 references of an independent review found.
Read the write-up →recolo.
2026– · Results committed
Designed memory for LLM agents (episodic and semantic stores with decay, salience and consolidation) and fixed the LongMemEval protocol before scoring. Plain retrieval over the same memories won (0.73 against 0.32 as shipped), so ablations and four more pre-registered protocols traced why and tested compaction. A redesign around what users add, revise and retract then beat both plain retrieval and rolling compaction on Memora at the same token budget (0.28 against 0.18).
Read the write-up →noul.
2026– · Results committed (find only)
Local code search that keeps code on the machine: BM25 and an embedding model shortlist candidates and a cross-encoder reranks them. On 35 labeled queries across two codebases, every single-file answer ranks in the top five at a tenth of brute-force cost; a strict every-correct-file metric showed only 4 of 9 multi-file answers complete, and the default stays unchanged until a fix holds on a held-out codebase.
Read the write-up →menhir.
2026– · Implemented
Co-owned training platform for athletes and coaches, heading into open beta, built under heurology with a coaching partner who owns the training content and engine logic. Designed and built the application from the ground up (screens, sign-in, data layer, AI wiring), now working alongside an associate engineer. A multi-provider LLM pipeline generates programs on top of 13 deterministic methodology engines, and every program is validated before commit. Invite-based accounts, tiered billing and an offline PWA; about 129K lines of code, 92 test files, 1,954 tests.
See the work →mara.
2026– · Shipped
Designed and built, under heurology, a multi-tenant pricing API for the moving industry, in production as a client platform's pricing service: one survey endpoint estimates weight from job-site photos, routes the move, recommends a crew and returns priced line items. Per-tenant rates and hashed, scoped API keys; 10 test modules.