Skip to content

Back2You

Intelligent lost-and-found — multi-tenant, white-label deployments

Category: Product & client work

Visit the live site(opens in a new tab)

  • Built with:
  • Next.js 16
  • React 19
  • TypeScript
  • Tailwind 4
  • Prisma 7
  • PostgreSQL 17 + pgvector
  • Redis
  • Voyage AI
  • S3/MinIO
  • MapLibre
  • Capacitor 8
  • Meta Graph API
  • Playwright
  • Railway
A lost-and-found search for the word backpack returning two confidence-ranked matches, each with its photo withheld behind a photo-hidden-until-verified placeholder.

What it is

Staff log found items with photos, location and metadata. Owners find them by typing a description or uploading a photo — matched with multimodal AI embeddings and vector similarity. QR “trackers” pre-tag ownership so a tagged item notifies its owner the moment it is logged, and a Messenger/WhatsApp bot lets anyone search with nothing installed.

Value

  • Turns a shoebox into a searchable inventory with a full claim audit trail.
  • Owners self-serve from web, mobile, kiosk or chat — no front-desk queue.
  • Privacy by design: photos stay hidden until staff verify identity in person.
  • One config file provisions a new institution — branding, roster, isolation.

Technology highlights

  • Multimodal AI search — Voyage multimodal-3.5 embeddings in Postgres + pgvector (HNSW): one vector space for text and image queries, guarded by a recall@k / MRR eval that fails the build on regression.
  • Server-enforced multi-tenancy — every read and write passes a Next.js 16 Data Access Layer that checks the role and injects tenantId. Middleware is never the security boundary.

More projects