Search clusters, not faces
Matching a selfie against a few hundred person clusters instead of every detected face makes search instant and also powers a 'browse by person' view.
Face-recognition photo delivery: guests take one selfie and get their own photos
Photographers upload thousands of event photos; guests scan a QR code, take a selfie, and instantly see every photo they appear in.
India
The challenge
What needed solving.
A wedding can produce 5,000+ photos across several functions. Guests do not want to scroll through all of them, and matching a selfie against tens of thousands of faces had to be instant on a phone.
What we built- Photographer studio and guest gallery (Next.js) with a Fastify API
- Direct-to-storage uploads with presigned URLs
- Ingest worker producing AVIF/WebP derivatives, thumbhash placeholders, and face embeddings
- Face clustering into people with union-find over a similarity graph
- Selfie search against cluster centroids stored in Postgres with pgvector
- Automatic event segmentation from photo capture time
Pluggable face provider
A provider interface lets the same pipeline run locally for testing or against a managed recognition service in production without code changes.
The result
Guests find their photos from a single selfie, and photographers deliver a whole event without manual sorting.
Related work
Similar projects.
Defense Net Solutions
The company website for Defense Net Solutions, plus two security products now run under their brand: a security posture scanner and a zero-knowledge credential vault with a Chrome extension.
Read the case studyForex brokerage
Marketing site, client dashboard, and admin panel in one Next.js app, connected to MetaTrader 5 through a dedicated bridge service.
Read the case studyRetail property collections team
A stakeholder-ready product prototype that ranks rent decisions by tenant sales and affordability, with evidence, approvals, and an audit trail.
Read the case study