I like figuring out how things actually work. Not the surface version, but the part underneath where you can see why an algorithm behaves the way it does. The fastest way I've found to get there is to build the thing myself, so a lot of what I make are from-scratch implementations of ideas I wanted to genuinely understand rather than just import.
I'm drawn to machine learning and reinforcement learning. I'm especially interested in how agents learn from trial and error, and in evolutionary methods, which I find genuinely fascinating in their own right.
When I run into something I don't know, I've learned to treat that as the interesting part rather than a wall. I'd rather be a little out of my depth and come out knowing the thing than stay where it's comfortable.
- SimpleSimpleNEAT — NEAT written from scratch: neural networks that evolve their own structure to solve control tasks. The project that got me hooked on evolutionary algorithms.
- transformer-gnn-layer — a graph-convolution layer dropped into a LLaMA decoder stack, to see what happens when you mix a transformer with a graph neural network.
- DocSeg2 and my bachelor thesis — document AI: pulling structure out of messy scanned pages and technical drawings with detection models and fine-tuned vision-language models.
- sNeuron-TST — worked inside a research codebase for a text-style-transfer paper and upstreamed fixes so it runs on current libraries.
- finaler-Beleg-2022 and a small x86 operating system — C++ from the ground up, down to a hand-written linked list and a kernel with its own keyboard and screen drivers.
- Race_history and Pic Roulette — full apps I built end to end, the second a multiplayer game that's live on the Play Store.