Operations
Reunitar is an offline-first lost-and-found operations app for large conventions. Staff run checkroom and lost-and-found counters — item intake, release, claims, disposition, missing-item reports, and person-case escalation — across multiple conventions, with role-scoped access and post-event reporting. I built it full-stack on Next.js and Prisma, including a local-first RxDB sync layer so staff keep working when venue Wi-Fi drops, and a semantic search layer (MongoDB Atlas vector + hybrid search) that matches free-text descriptions of lost and found items.
Context
Lost-and-found at a big convention is a logistics problem: thousands of items, rotating volunteer staff, anxious attendees, and notoriously unreliable venue Wi-Fi. Reunitar gives staff a single operations workspace that keeps running offline, matches found items to missing reports by meaning rather than exact keywords, and gives supervisors coverage alerts and post-event reporting.
Constraints
Role & Team
Process
01
I built the ops UI on RxDB (IndexedDB) with custom pull/push sync endpoints. Item intake and person-cases are fully readable and writable offline; changes replicate when connectivity returns. Forms try the server first, then silently fall back to local creation on network failure.
02
I layered in conflict resolution (automatic server-wins plus a human-in-the-loop merge UI) and a signed session-cache JWT so authentication survives a 24-hour database outage — because at a live event, "the server is down" can't mean "stop working."
03
Lost items rarely match found items by exact words ("black hoodie" vs "dark sweatshirt"), so I added vector and hybrid search with a keyword fallback. I then migrated the whole embedding pipeline from PostgreSQL + pgvector + Redis + BullMQ + OpenAI to MongoDB Atlas Automated Embeddings (Voyage AI) — Atlas embeds on write, which removed four moving parts from the backend.
Outcome
Staff roles with scoped, offline-capable access
Search strategies: vector, hybrid, text fallback
Backend services removed in the Atlas migration
Lessons
Designing for genuine offline means distrusting navigator.onLine and treating sync, conflicts, and even auth as first-class offline problems — and managed embeddings (Atlas + Voyage AI) bought that resilience back by removing a whole self-run OpenAI + Redis + BullMQ pipeline.
