Building a decoder-only transformer in pure PyTorch: tokenization, embeddings, causal multi-head attention, residual streams and layer norm, trained on Apple MPS to own every gradient that flows.

Collected 60,000 abstracts across eight fields from the OpenAlex API, about 9.45 million words, and turned them into a training corpus and a semantic-search engine with embeddings and >_FAISS.

Taught a small car to drive itself: HSV lane segmentation, lane-offset steering and an ultrasonic safety stop on a Raspberry Pi running ROS 2 Jazzy.

Prototyping retrieval-augmented and agentic workflows with LangChain, LangGraph and local models: small systems that read before they answer.

Placeholder: a clean FastAPI backend for clinical workflows: typed schemas, auth, and tests that run before anything ships.

Studying computer science in Berlin with a focus on deep learning, and spending the gaps between lectures training models that have no business fitting on a laptop.