Mathieu Tellene

Fourteen aircraft boarding methods race on the identical passenger manifest, so any difference is the ordering and not the luck of the draw. The model contradicted four of my own hypotheses — each is documented with the measurement that overruled it. One HTML file, no dependencies and no build step, with a 3D renderer written from scratch: perspective projection and painter's algorithm on a 2D canvas.

Workforce planning end to end on real open data: forecast the contacts, size the floor, build a roster that obeys the law, and put a price on it. Runs on 3.17M rows of NYC 311 open data, sizes each channel with the model that actually applies to it — Erlang C for voice, a concurrency model for chat, backlog conservation for tickets, because email is not a queue — and solves the week with CP-SAT under Spanish working-time law. The whole pipeline is also ported to JavaScript, so you can move the headcount on the page and watch the distribution matrix rebuild.

Music recommenders are open loop: they know what was played, never whether the floor emptied. Kadenz closes that loop with computer vision — pose estimation and optical flow measuring crowd response in real time — and ranks the same catalogue both ways at once, so the premise is testable instead of asserted. Runs on a laptop CPU; the whole loop also runs client-side in a browser.

A city you have never operated in has no ride data. It does have satellite imagery. A small CNN reads demand off Sentinel-2 chips, trained on New York alone and then scored cold on Chicago — 77 community areas it had never encountered. The honest result is the interesting one: 0.88 rank correlation but only +0.23 R², so the ordering transfers and the level does not. It will tell you which neighbourhood outranks which; it will be off by 2.65× on how much. Click any zone on the map and watch its chip pass through the network block by block.