Swapzy

Challenge

Face-swap video is generative media: each job needs a GPU, runs seconds to minutes, can fail, and costs real money per second of compute. Built naively, it burns budget idling or leaves users waiting on a cold queue. It had to feel instant to a first-time visitor and stay economically sane at near-zero traffic.

Process

The pipeline is fully asynchronous. An upload fires an S3 event to SQS, a GPU worker runs FaceFusion (ComfyUI is a second engine for stylized swaps), and CloudFront delivers the result. The API runs on a multi-AZ Auto Scaling Group. The Spot GPU fleet scales on queue depth and keeps at least one warm worker, so a new user’s first job never hits a cold start; I made that cost-versus-latency call with the math written down.

Security and reliability are first-class. GitHub Actions deploys via federated OIDC (no stored keys) and tag-scoped SSM. The runtime database user is DML-only, separate from the migration owner. An append-only, hash-chained audit log anchors hourly to S3 Object Lock in WORM mode. There are 21+ CloudWatch alarms, and a chaos-drill campaign I designed validated the system, including a database restore in 7 minutes 25 seconds against a 30-minute target. A Stripe-billed Pro tier and an admin-gated entitlement system are built and verified end-to-end on production.

Result

Live in an audience-scoped soft launch (friends-and-family beta) at swapzy.co. The generative pipeline, payments, and reliability engineering are built and proven; I keep it in restricted beta by choice, closed to the public. It is the only portfolio project with a GPU generative-media workload at its core, with the architecture and economics that implies.