scratchnet ★
Tiny autograd + MLP from scratch in 700 LOC of Python.
Built to teach myself the maths. Trains MNIST in under two minutes on CPU. The tape-based gradient bookkeeping is what most surprised me.
Software engineer · ML & web platforms
Building tools that compound — small interfaces over deep systems.
Sara is a software engineer who likes the boring half of engineering: clean APIs, fast tests, and observability you can act on. She's worked on ML platforms, payments infra, and the occasional indie game. Currently exploring what AI assistants should never do for you.
Tiny autograd + MLP from scratch in 700 LOC of Python.
Built to teach myself the maths. Trains MNIST in under two minutes on CPU. The tape-based gradient bookkeeping is what most surprised me.
Kafka log inspector for humans — one keystroke per task.
TUI for browsing topics, filtering by header, and replaying ranges. Ergonomics over feature count.
Building the eval platform — dataset versioning, run diffing, and a CLI good enough that researchers stopped writing their own scripts.
Owned the merchant-onboarding flow end-to-end; cut KYC median from 6h to 38m via async webhooks + retry policy.
Senior thesis on robust active-learning under label noise. TA'd Algorithms II for two semesters.