Stanislav Tsepa

Stanislav Tsepa

Senior ML Engineer & Researcher
London

Any sufficiently advanced matrix multiplication is indistinguishable from magic.

Projects

autoevolve★22Aug 2026

I made an autoresearch loop where the benchmark is self-play — the agent mutates its strategy overnight, plays it against its earlier versions, and keeps an honest Elo rating to pick the winner. Tested on a handful of AI strategy competitions

I wrote a DSL layer that deterministically builds any factory from its inputs and outputs, using chip-design algorithms to solve routing and congestion

manim-ai-agent★12Aug 2026

I made an agent write, render and critique its own 3Blue1Brown-style maths videos until they're watchable

I masked 16×16 holes in factory blueprints and trained a transformer to fill them back in

mafia-battle★2Feb 2026

I made LLM agents play Mafia, the classic social-deduction game, against each other — to measure how well different models deceive, and how well they catch each other at it. Every game is logged and replayable in the browser

Writing & talks

Talk at Data Sanity — using LLM-as-judge evaluation and automated prompt optimisation to build better models

First-author paper at IEEE ICDM, with Maxim Panov

Career

A year and a half at Meta working on end-to-end AI support — LLM agents that take a user's problem from first message to resolution instead of handing it to a person. SFT→RL post-training loops, reward modelling, and the LLM-as-judge eval layer underneath. Two years at Booking.com on transformer user-intent models serving millions of predictions a day. Three years at Yandex Self-Driving on the prediction and planning layers of the L4 stack, some of which showed up at CES 2020.

MSc Data Science at Skoltech + MIPT. Full CV →