Clearcut

The problem

remove.bg is closing its standalone product on 1 December 2026. Its users need a fast, cheap replacement with an API, and they need it before that date.

What I built

A web editor and a paid API. It has accounts, hashed API keys, OpenAPI docs and an atomic credit ledger with reservations and refunds. Billing runs on Polar, with webhook reconciliation that handles partial refunds. Free and paid users go into separate queues. It launched with a status page, SEO and comparison pages, llms.txt, and a cost-monitor Worker that emails me when GPU spend passes $2 an hour or $10 a day.

How the AI works

• Background removal: BiRefNet Lite (fast) and BiRefNet (fine) on serverless GPUs. Revisions are pinned, containers restore from memory snapshots, and requests are batched dynamically.
• Smart Brush: SlimSAM selects an object from a painted stroke. I rank mask candidates by how much they overlap the stroke, not by size. A stroke on a car’s bonnet now selects 5.8% of the frame instead of the whole car at 47.8%.
• Magic Background: a generated scene behind the subject. The subject is never repainted, and a deterministic contact shadow keeps it from floating.

What I measured

• $0.63 per 1,000 fast removals, from a 1,000-image cost audit.
• Latency from about 7 s to about 1.7 s by returning only the alpha mask and keeping the user’s original upload.
• 43 GPU snapshot restores measured (median 3.48 s, worst 29 s). On that evidence I dropped a sub-second cold-start goal instead of promising it.
• An ONNX export of BiRefNet (opset 22, custom DeformConv) matched PyTorch at IoU above 0.9995. I rejected it because it wasn’t faster.

Why it matters for clients

It is the whole job in one small project: model choice, GPU cost engineering, billing, abuse limits, SEO, monitoring and launch, all done in under a week.