Presidential AI Challenge
Crowd chokepoint detection using the thigmonastic response of Mimosa pudica — a living sensor instead of a camera array.
I build instruments for things that are hard to measure.
Ten systems, across financial literacy, rare-disease target triage, drug-discovery lead mining, competition robotics, and a crowd sensor made of a plant. Each one takes a signal nobody had a clean read on and turns it into something you can act on — and argue with.
Repaints this page as one of 25 complete design systems.
Every upstream call has to fail softly. Partial data renders, missing data is omitted, and nothing is invented to fill the gap. When raremp's hero can't reach Open Targets it falls back to a generic helix and drops the worked example rather than showing a number it can't source.
A ranking nobody can inspect is a ranking nobody should trust. raremp exposes its weights and itemises every contribution on a first-class surface. A competitor can copy the data sources; it can't copy a score the user is allowed to disagree with.
Demo mode is a way to avoid the hard part. Divendo runs on live Polygon quotes and real Supabase accounts; raremp renders from live upstreams at request time. Mocks survive as fallbacks, never as the thing being shown.
PocketScout excludes DrugBank, ChEMBL and Pharos because their licences don't permit what the product needs, and rejected fpocket for being GPL-3.0 with no Windows build. The licence policy isn't legal housekeeping around the product — it shaped what the product is.
A market that behaves like the real one, inside a structure a teacher can actually run a class on.
Trading simulators have no teacher surface. Finance curricula have no live market. Divendo is both halves: paper trades against real Polygon quotes, wrapped in curriculum, a gradebook, submissions, a live classroom, and class analytics. Students include under-13 users, which makes COPPA parental consent a legal requirement rather than a feature — an under-13 signup cannot be approved until a consent email is actually delivered.
The sensor is a plant.
Mimosa pudica folds its leaves when it is disturbed. CrowdLeaf reads that thigmonastic response as a signal and turns it into a map of where a crowd is jamming — chokepoints found by a living sensor rather than a camera array. It won the Presidential AI Challenge for the District of Columbia.
A ranking you are allowed to argue with.
A researcher arrives with a disease name and leaves with a defensible shortlist of proteins worth pursuing. Twelve upstream databases are collapsed into one traceable path — disease, then target, then structure and variants. The scoring weights are exposed and adjustable, and every contribution to a score is itemised on a first-class surface rather than hidden in a tooltip. A competitor can copy the data sources; it cannot copy a ranking the user can disagree with.
Druggability × novelty. One multiplication is the entire product.
A gorgeous binding pocket on EGFR is worth nothing — 82 drugs already target it. A merely good pocket on a transporter with 166 papers and zero chemical probes is the whole business. PocketScout mines the AlphaFold Database for well-formed pockets on proteins nobody has characterised and ranks them as leads. Pocket-finders ignore novelty; novelty databases ignore druggability. Neither half is the product.
Wander well.
An AI event-discovery app that has to justify what it recommends. The anti-cheat economy is the interesting part: pressing “I'm going” is a free RSVP that grants nothing. Experience points come only from photo-verified attendance and one review per event, enforced at the database with unique constraints and triggers rather than in the client where it could be gamed.
A field you can draw a route on, at 480×240, on a robot.
Competition code for VEX V5RC team Lucky Cats, running PROS 4 with a LemLib chassis and an LVGL touchscreen UI. The Design screen is the part worth showing: tap the field to drop waypoints, build a route by hand, and run it without leaving the screen. The whole selection — alliance, route, start pose, and the custom route itself — survives a reboot as plain text on the SD card, safe to edit from a laptop. With no card inserted it falls back to defaults instead of failing.
The unit of value is a finding nobody knew to look for.
Most tools for organisational waste require somebody to file a report about it — which means the waste nobody notices stays invisible. PHANTUM scans an organisation's own working artefacts and surfaces the effort being duplicated, abandoned, or ignored, without anyone having to ask.
An arena, not a marketplace.
Challenges are the unit, not gigs. A challenge is posted, creators submit work, entries are judged, results are public — competitive framing around creative work, carrying real payments and file uploads on the critical path. Deliberately built on a different stack from everything else here: Pages Router, JavaScript, Prisma.
The site is part of the argument that the team is serious.
A public site for a VEX Robotics team, aimed at sponsors and recruits who arrive knowing nothing. A cinematic intro leads with the team name, then the season's real match history renders to canvas as a score series — nineteen matches, actual results, not a decorative chart.
Takes, groups, and a vertical feed.
A React Native application built on Expo Router — a full-height paging feed of featured “takes”, plus groups, a create flow, and profiles. Early: the architecture is in place and the surfaces are scaffolded.
Pick the line that fits and it goes straight to the right inbox. I read everything and reply to anything specific.