stuff I actually
built & broke.
No "synergistic enterprise solutions," no fabricated 98.4% benchmarks. Just real side projects, a bit of research, and the occasional thing that only works on my machine. Honest status on every one.
A TypeScript tool + GitHub Action that bidirectionally syncs my Notion database with this Jekyll blog. Runs on an 18-hour cron, preserves front-matter, and quietly deletes the posts I unpublish.
Research-internship work at the Physical Research Laboratory: a neural network that folds hyperbolic equations into its architecture. Cut compute time by ~37% over the baseline while holding accuracy steady.
My ongoing playground for evaluating multi-agent LLM workflows — small eval harnesses with DeepEval, poking at exactly where agents fall apart, and trying to measure "did it actually work" instead of going on vibes.