Instruments, not oracles.
I build systems that report what they don’t know, and tell you when the result they just gave you is worth nothing.
Solved in your browser at 30 frames per second. Click or drag the field to drive it.
About
Hi, I am Vincent, a Computer Science student at ETH Zurich, working across probability, machine learning and systems. The thread through all of it is calibration: abstention when the evidence is thin, correction for the size of the search, and honest reporting of a result that went against me. Away from the screen I am currently learning to bake bread, getting into running, and spending a lot of my free time hiking.
Before ETH I received multiple prizes at state level in Jugend Forscht in Biology. I was also a finalist in the 42nd Bundeswettbewerb Informatik and placed in the top ten in German competitive programming. School had no computer science course, so that part was all self-taught. I play the violin on a merit-based instrument loan from Karl Montag and grew up speaking Chinese, English and German. As a teenager I spent five years training about thirty bonsai off nothing but YouTube tutorials, so I know an unreasonable amount about them.
Projects
Open-ended work in progress.
An experimental artificial memory system that reconstructs information from incomplete clues (fragments, context, associations) rather than exact queries.
0.943 Recall@5, associative
0.886 Recall@5, vector search
40 queries over 44 interconnected memories
Finite-difference PDE solvers written from scratch, an interactive lab to drive them, and a measurement of how much of a GPU they actually use. Writing a 5-point stencil the readable way costs about 33 passes over memory where two would do; a hand-written fused Triton kernel computing the identical update gets that back.
93% of achievable bandwidth, fused Triton kernel
7% of achievable bandwidth, eager PyTorch
4.2–12.9× faster, measured on an RTX 2080 Ti
Apps
Software that runs: deployed, or installable and open source.
A self-hosted web app that replaces physical poker chips for in-person Texas Hold'em, tracking stacks, bets, pots and side pots from players' phones over a shared room code. It never sees a card: it leans on the table to declare the winner at showdown, the way players push chips across felt. Core functionality is complete; licensing and polish are still in progress.
A local post-match analytics tool for Counter-Strike 2 demos: drop in a .dem and get a per-player scoreboard, a round-by-round event timeline, and kill and death positions plotted on the real map radar. Every statistic is computed independently of the parser and has a published definition, checked against both the parser's own aggregation and a third-party scoreboard on real matches; a metric the demo cannot support renders as an em dash with the reason rather than as a zero. Nothing is uploaded anywhere: the server binds to localhost and the demos stay on your machine.
A camera-driven multiplayer party game, and the boilerplate for building one. Every player's phone posts camera stills over HTTP to a Python backend that holds the authoritative game state and pushes it back over WebSocket. Two channels on purpose, so a slow upload can never stall the state the clients render. It ships two modes: shared rounds, and an asynchronous race in which everyone walks the same server-side sequence of prompts that no client is ever allowed to read ahead. The frame processing is a deliberate stub; the plumbing and the game rules are real.
Socials