Gaspard Moulinier

A phone-first autonomous agent that runs real web workflows on a persistent cloud browser, so work gets handed off from a phone with no computer in the loop. I wrote the browser runtime in-house rather than renting it, and every action passes a deterministic policy gate that marks it allowed, confirmation-required or blocked before it runs. In build.

An iOS journaling app for walks. It records the route you took, the voice notes you left and everything you did along the way, so you can go back and relive it. Everything stays on the device, with no personal-data servers. Over 3,000 users. Swift, MapKit and Core Location.

A reverse marketplace for collectible cars. Buyers post the car they are hunting, finders track it down and earn the fee. Next.js and Supabase on top of a Python scraper indexing 2,100+ listings and 16,000+ photos across seven US sources. 212 cars found for buyers under its previous name. In build.

A cross-asset hedging engine built in 36 hours with a five-person team at the Paris Builds hackathon. It puts an equities book and a prediction-market book behind one hedge-ratio engine, pricing event risk as an explicit, quotable cost of protection, including the quantisation error that dominates basis risk on small books.

A sole-authored working paper on SSRN. A published claim of persistent, structurally protected arbitrage between the two largest prediction-market venues is tested on executable quotes and does not survive. The apparent mispricings are timing artefacts: deliberately shifting one venue's quotes against the other makes the edge grow monotonically. The binding cost on these venues is the bid-ask spread, not the trading fee, by roughly a factor of six.

Classifies trend-continuation against mean-reverting states on NQ futures from Level 2 quote data rather than from price bars. Quote-level pressure accumulates in a volume-clocked reservoir, signed by trailing trade displacement and normalised by rolling realised volatility, calibrated strictly causally. The baseline chain is frozen, and any improvement has to beat it under purged walk-forward tests with forward net-of-cost labels. Classification research, no live edge claimed.

An autonomous loop that generates, pre-registers, tests and kills high-frequency theses between iterations with no human input, pricing queue position, spread crossing, fees and latency. Fourteen documented kills and one structural finding, which became the paper above.

A decision pipeline run over the OEIS corpus of 399,049 entries, 45,377 of which carry an unproved conjectural line. 17,368 entries settled across two runs, the most recent covering 10,372 separate conjectural statements. Anyone can check it: verify.py rebuilds every result of that run from scratch in exact integer arithmetic and passes 5,920 of 5,920, and a withdrawal is published whenever a premise fails. With Adrian Perez Fontelles.