← Back to Labs
UnGovr Reunite

UnGovr Reunite

Building

Overview

Lost items don't respect jurisdiction lines. A wallet lost on a county bus might be turned in at a city transit office, a state park, or a neighboring municipality's police station. Today there is no way for the person who lost it to know where to look — or for the agency that found it to reach the right person.

UnGovr Reunite is a cross-jurisdiction matching system with two entry points: Found (photo-first — a finder posts what they found) and Lost (text-first — a seeker describes what they lost). AI classifies items, generates semantic embeddings, and matches across boundaries. When a match is found, a privacy-preserving relay thread lets both parties communicate without sharing personal contact information.

AI Classification

Every found item is analyzed by Claude Haiku's vision model. For photos, the AI detects the object, extracts color, brand, and distinguishing attributes, and flags sensitive content like government IDs, medical devices, or visible serial numbers. For lost reports, the AI classifies the text description and extracts product URLs if provided.

Brand names (100+ patterns) are automatically stripped from public display to prevent targeted theft — a finder's photo showing an Apple logo is classified internally but the public listing says "phone" not "iPhone."

Privacy-First Design

UnGovr Reunite treats every piece of personal data as potentially dangerous. The system is designed so that even a full database breach would not expose enough to target anyone.

  • Phone numbers stored as SHA-256 hashes only — originals are never persisted
  • Location shown as coarse geohash (±50m) — exact coordinates never exposed
  • Time shown in buckets (<24h, 1–7d, 8–30d, >30d) — not exact timestamps
  • Photos restricted to government staff and verified finders — not publicly browsable
  • Brands stripped from public display to prevent "treasure hunting"
  • VoIP numbers blocked at verification — mobile carrier numbers only

Visibility Tiers

Each found item is assigned a visibility tier based on sensitivity. Government IDs and medical items default to the most restrictive tier; everyday items like clothing default to public browse.

  • Tier A — most restrictive: gov/staff only (IDs, medical, sensitive documents)
  • Tier B — restricted: verified users only (phones, laptops, jewelry)
  • Tier C — standard: public browse, photo restricted (most items)
  • Tier D — open: fully public including photo (low-value, non-sensitive)

Item Categories

The system classifies items into 15 categories, each with default sensitivity levels and retention periods. Sensitive items (IDs, medical devices) are retained for 14–60 days; normal items (clothing, bags) for 120–180 days.

  • High sensitivity — government IDs, medical devices, prescription medications
  • High value — phones, laptops, tablets, jewelry, cameras
  • Standard — bags/backpacks, wallets/purses, clothing, keys, glasses/sunglasses, headphones, books, water bottles, umbrellas, other

Claims System

When a seeker believes a found item is theirs, they submit a claim with tiered verification. The finder reviews the evidence and decides whether to release the item. This prevents false claims while keeping the process lightweight.

  • Receipt or proof of purchase — strongest evidence, uploaded as image
  • Product listing link — URL to the exact item purchased (extracted by AI if provided)
  • Detailed description — distinguishing marks, contents, or features only the owner would know

Why Government Infrastructure Matters

Existing lost and found apps are fragmented across commercial platforms. Government agencies — police departments, transit authorities, parks — already collect found items but rarely share data across agency lines. UnGovr has the entity map and the interagency relationships to solve the cross-jurisdiction problem that commercial apps cannot.

The privacy-preserving relay is essential: people reporting lost items to a government-adjacent platform should not have to share personal contact info with strangers. Relay messaging (similar to rental car apps and marketplace platforms) solves this while keeping the exchange safe for both parties.

Build Status

Status: BUILDING

Found item posting
done
Lost item reporting
done
AI classification
done
Matching engine
done
Privacy controls
done
Relay messaging
done
Claims system
done
Admin dashboard
done
Map interface
done
Event mode
planned
Match notifications
planned
Reward processing
planned

How It Works

Posting a Found Item

# Finder uploads a photo — AI handles the rest
→ upload photo attached (EXIF stripped, faces detected)
→ classify Claude Haiku vision → backpack, black, outdoor gear
→ strip brand name removed from public display
→ locate coarse geohash: Ventura Fairgrounds area (±50m)
→ embed semantic vector generated (pgvector)
→ tier visibility C — public browse, photo gov-only

# Item is live on the map and searchable
✓ Posted — retention: 120 days

Reporting a Lost Item

# Seeker describes what they lost — text-first
→ describe "Lost black backpack at Ventura Fairgrounds, Sat Oct 12"
→ classify Claude Haiku text → backpack, outdoor gear category
→ embed semantic vector generated
→ scope search radius: city → county → region

# Immediate scan against all found items
→ scan 3 candidates above 0.6 threshold

AI Matching

# Composite scoring: 4 weighted signals
→ semantic cosine similarity (pgvector) — 50% weight
→ category exact=1.0, related=0.5, different=0.0 — 20% weight
→ location geohash prefix overlap — 20% weight
→ temporal decay over 30 days — 10% weight

# Result
→ match 0.84 — "Black backpack, Oct 13, Channel Islands Harbor"
→ action above 0.8 threshold → relay thread opened

Relay Messaging

# Privacy-preserving communication
→ thread relay opened between finder and seeker
→ relay neither party sees the other's phone number
→ lang auto-translate between EN and ES-419
→ consent both parties must opt in to share direct contact

# Claim verification
→ claim seeker submits proof: receipt, purchase link, or description
→ verify finder reviews claim evidence
✓ Matched — item marked as returned

Tech Stack

Claude Haiku (vision) pgvector PostGIS FastAPI PostgreSQL SMS relay Geohashing Python