Hi, I'm Younes

Two connected parts. A weekly risk model for every player in Europe's top five leagues: calibrated logistic regression on 1.29M player-weeks of public data, tested once on a season it had never seen. And a live match-day tool for bench staff that compares each player with his own normal, runs offline at the ground, and was built and tested end to end on synthetic matches.

Weekly model: top 10% risk group injured 6× as often as the bottom 10%.

Cost-aware backtesting on Binance spot data: fees and slippage on every fill, walk-forward validation, out-of-sample results only, bootstrap confidence intervals.

Where the edge did not survive costs, the report says so.

A small CNN reads handwritten sudoku grids, a SAT solver solves them, and MaxSAT repairs the digits the network misread.

Fully correct grids: from 75% to 99% on 1,000 test puzzles.

A multi-agent LLM system for mathematical proofs: Claude agents plan, write and repair proofs, and an independent Lean 4 checker decides what is accepted.

Nothing counts as proved unless Lean verifies it.

SAT encoders and small stand-alone verifiers that re-check every bound in my thesis, using nothing beyond standard Python.

Every thesis bound reproduced in seconds, on every push.

An interactive companion to the thesis: paint colourings of 2D and 3D grids and watch every line get checked live in the browser.

The quickest way to see what the thesis is about.

A live multiplayer browser game used as a testbed for invisible-prompt defenses: hidden Unicode instructions aimed at AI tools (ASCII smuggling, emoji smuggling, Trojan Source) are stripped in Postgres before anyone reads them.

20 attack payloads in CI; the tests found 4 gaps, all closed; live site 29/29.