Vojtech Letal
Software engineer · Google DeepMind · Zurich
I work on Gemini's post-training — the part after the model knows how to talk, where it learns what to do and what not to fall for. Mostly alignment and adversarial robustness: red-teaming it, breaking it, and building the data and training that make the breaks stop working.
Work
Software Engineer · Google DeepMind
Core modelling and post-training for Gemini.
- Built training data that raised the cost of a prompt injection against Gemini roughly 10×.
- Shipped targeted response modification (SFT/RL) — ~90% user acceptance.
- Productionised automated red-teaming that finds prompt-injection holes before anyone else does.
- Adversarial image attacks against Gemini's vision stack.
Machine Learning Engineer · Google
Bard. Making a large language model actually follow what a person meant, not just what they typed.
Staff ML Engineer / Scientist, AI · Avast
Machine learning against people who are actively trying to defeat it. Prague.
Senior ML Engineer · Blindspot Solutions
Custom optimisation and game-theoretic systems for corporates and startups. Prague.
Things
SKOKAN
A one-button platform runner for an ESP32 arcade box I built with my 8-year-old. The same C++ runs on the board, on macOS and in your browser — click to play.
OBEGRÄNSAD over Bluetooth
New firmware for IKEA's 16×16 LED wall lamp: it runs without WiFi, pairs with the browser over Bluetooth and takes firmware updates the same way. Comes with a small procedural dude, trapped in his box.