Lumina

Lumina turns raw photo storage into an intelligence layer. It detects faces, clusters identities, and lets people instantly find themselves in large, shared photo collections, the way Google Photos does, but without locking users into a proprietary system or taking control of their data.

What it does

• Collaborative Events. A host shares a QR code, guests upload photos, and instead of scrolling through thousands of strangers a guest just takes a selfie. A vector search finds every photo they appear in within that event.
• Bring Your Own Storage (BYOS). Each album defines its own backend (Local, Cloudflare R2, or S3). Lumina holds the intelligence (embeddings, metadata) but never the high-res originals.
• Asynchronous AI processing. Uploads return immediately; the heavy work runs in a background pipeline.

How it’s built

The system is a distributed pipeline that keeps heavy computation off the request path:

• Bun + ElysiaJS API handles validation, auth, and usage quotas.
• BullMQ worker orchestrates the pipeline (image optimization with Sharp, 64-bit dHash perceptual de-duplication, then embedding extraction) with retries, idempotency, and backpressure.
• FastAPI ML service keeps the InsightFace buffalo_l model warm in memory, extracting 512-dimensional face embeddings. Moving from a per-image Python subprocess to a warm service dropped latency from ~400–800ms to ~40–100ms per image.

The interesting decision

Face similarity search runs on plain PostgreSQL without pgvector: embeddings are stored as Float[] arrays and cosine distance is computed on the fly with unnest() and standard SQL. That keeps Lumina portable to any Postgres instance (Neon, Supabase, self-hosted) with zero extensions. Because searches are scoped per event or album, the O(n) scan stays small and fast.

For the full architecture, the pgvector-free similarity query, and the reliability/backpressure design, read the deep dive linked above.