Everything I'm building — research, tools, and interactive explainers. Each one belongs to a track; open a project to see where it fits.
Six months of coffee, measured.
I photographed 39 empty coffee bags, had LLM agents transcribe every label, and geocoded each journey — 360,761 km from farm, to roastery, to a cup in London. A 3D globe replays all six months, bag by bag.
Every black hole begins as a breath of dust.
An interactive, scroll-driven story of a star's whole life — gravitational collapse, the ten-million-year fusion truce, an iron core that can't burn, the supernova, and the three endings decided by mass alone.
A conversational book recommender.
Tell it a feeling, a question, a mood — get back two or three books, chosen by a librarian-friend that builds a picture of your taste over time. Four personas, a prose taste portrait, and a theme constellation.
I gave Gemma 4 my Instagram reels.
A two-stage multimodal pipeline that watches your reels, finds the through-line in your content, and tells you who you're really making things for — strategy grounded in actual video, not generic advice.
Can an LLM read behavior the way a recommender engineer reads a session?
A six-part arc on LLMs in recommendation — from prompt-based preference prediction and scaling laws to synthetic reasoning data, GRPO, generative recommenders, and a repo-backed LLM cross-encoder.