My toolbox

Finding the camera in a brain scan

At Harvard Medical School, I used neural radiance fields to study how a surgeon’s camera view could line up with brain scans. I optimised camera poses through backpropagation and compared loss functions, including mutual information, for accuracy and convergence.

Learning to play, then searching ahead

My Pokémon agent combines behaviour cloning, PPO, and self-play with deck-specific LoRA specialists. At match time, it uses PUCT tree search over possible opponent decks. All within two CPUs, with no GPU or internet.

Built with PyTorch, nerfstudio, Hugging Face, Unsloth, and MLflow.

From an issue to a code change

At Apple, I designed and shipped a system that takes software issues through to code changes ready for review. That includes the tool use, execution traces, and fallbacks around the models. I also train smaller models for the work, using supervised fine-tuning and self-distillation policy optimisation (SDPO).

The data has to get there, too

At Expleo, I built pipelines processing more than a million daily data points for BMW’s autonomous-driving projects. I’ve also built web applications for legal and investment workflows.

React, Next.js, FastAPI, PostgreSQL, Spark, Docker, and Kubernetes. Also in the toolkit: AWS, Cloudflare Workers, Fly.io, and GitHub Actions.

I spent two years as president of TUM.ai, helping grow a community of 300+ members. We ran summits and hackathons, worked with industry partners, and started a research department.

I’ve coached startups at UnternehmerTUM Venture Labs, helped teach a two-week AI programme with KNUST in Ghana, and co-organised a defense hackathon with 400+ participants.

Mathematics

I studied Information Systems and Mathematics at TUM. Statistics, optimisation, information theory, and differential geometry are part of that background.

Languages

Python, TypeScript, SQL, Go, Java, Kotlin, C, and C++. With people: English, German, and Ukrainian.