pluckless

pluckless is the most ambitious thing in this list, and the one that carries the ML and infrastructure story.

The job sounds simple — turn a photo into a clean SVG — but doing it well means real models, not heuristics. A request fans out across a small pipeline: BiRefNet mattes the subject from its background, SAM 2 segments it into coherent regions, and a vectorizer turns those regions into layered, editable paths.

The architecture is the case study. Cloudflare Queues decouple upload from the expensive GPU step; R2 holds the originals and outputs; D1 tracks job state; and a Fly GPU worker does inference and reports back. It’s a practical pattern for putting heavyweight models behind a fast, cheap edge front door.

Private project — featured here as a writeup with screenshots, no public repo.