LostNet — The Neighborhood Lost & Found with Gravitational Resonance & Real-World Handovers

This is a submission for the Sanity Challenge, Path Two: Vibe-Code Something Strange

---

What I Built

LostNet is a real-time, map-first neighborhood Lost & Found network where lost items gravitationally pull toward their finders.

Set in a bustling city neighborhood (Indiranagar, Bengaluru), LostNet operates as zero-friction public infrastructure: no account, sign-up, or password required. If you drop your keys or find a phone at a bus stop, you can report it in 15 seconds directly from your phone.

A deterministic Gravitational Resonance Engine continuously evaluates spatial, temporal, and semantic similarity across four distinct pillars:

• Category Match (35%): Strict taxonomy alignment (Keys, Phone, Wallet, Pets, etc.)
• Geo Proximity (30%): Inverted exponential decay based on neighborhood walking radius
• Time Window (20%): Chronological plausibility (e.g., found after lost within a 60-hour window)
• Attribute & Description Similarity (15%): Color, material tags, and textual cosine matching

flowchart TD
    subgraph Engine["Gravitational Resonance Engine"]
        A["Item A: Lost"] --> M["4-Pillar Resonance Evaluator"]
        B["Item B: Found"] --> M
        M --> C1["Category Alignment (35%)"]
        M --> C2["Geo Walking Decay (30%)"]
        M --> C3["Temporal Window (20%)"]
        M --> C4["Attribute/Color Cosine (15%)"]
        C1 & C2 & C3 & C4 --> Score["Composite Resonance Score (0-100%)"]
    end
    Score -->|Score >= 55%| ResonanceLine["Visual Gravitational Vector on Map"]
    Score -->|Score >= 75%| HandoverTrigger["Match Dialog & Handover Activation"]

---

What makes it "Strange"?

Most lost-and-found software tries to build an airline ticketing desk on the web: multi-step account creation, complex ticket IDs, KYC verification, and bureaucratic hurdles.

LostNet models how humans actually behave in a city. When people find lost belongings, they do one of three things:

1. 🏛️ Drop it at known civic infrastructure (local police station reception, metro customer desk, café counter)
2. 🤲 Keep it safe while waiting for verified contact
3. 📍 Leave it undisturbed at the exact spot

LostNet embeds these real-world drop-offs directly into the content lifecycle, paired with an anonymous direct coordination chat so owners and finders can safely arrange pickups without exchanging personal phone numbers.

sequenceDiagram
    autonumber
    actor Owner as 👤 Item Owner
    actor Finder as 🔍 Item Finder
    participant Board as 🗺️ LostNet Map
    participant Sanity as ⚡ Sanity Content Lake
    participant Chat as 💬 Direct Handover Chat

    Finder->>Board: Pin found keys + Select "Dropped at Indiranagar Police Station"
    Board->>Sanity: Write lostFoundItem doc (Published)
    Owner->>Board: Report lost keys with blue lanyard
    Sanity-->>Board: Real-time Live Listener fires
    Note over Board: Gravitational Engine detects 88% resonance
    Board->>Owner: Display Resonance Vector on Map
    Owner->>Board: Click "Inspect Match & Coordinate Handover"
    Board->>Chat: Open Anonymous Handover Room
    Owner->>Chat: "I can come by the police station at 5 PM!"
    Finder->>Chat: "Desk officer has it tagged under Meher"
    Owner->>Board: Tap "Confirm Item Returned"
    Board->>Sanity: Create reunion document + Resolve items
    Board-->>Owner: 🎉 Confetti + City Reunions Ledger updated!

---

Demo

• 🌐 Live Application: https://lost-net-todt.vercel.app/
• 💻 GitHub Repository: https://github.com/ANSUJKMEHER/LostNet

Autonomous Guided Tour (Interactive In-App Demo)

Rather than forcing visitors to watch a pre-recorded clip or manually create dummy items, LostNet includes a built-in Auto-Demo Mode directly in the UI.

Clicking "Auto Demo" in the top navigation activates an automated 3-stage camera choreography:

1. Stage 1 (Resonance Animation): Automatically focuses the map on a high-resonance lost/found pair and animates the gravitational vector pulling them together.
2. Stage 2 (Inspection): Slides open the match proposal showing confidence breakdowns, distance, and the finder's physical handover location.
3. Stage 3 (Reunion Flow): Steps through the handover confirmation, fires celebratory particle physics, and logs the completed handover to the public ledger.

Screenshots

Interactive Resonance Map	4-Pillar Match Verification	City Reunions Ledger

---

Architecture & Real-World Handover Design

graph LR
    subgraph Client["Next.js 16 Web App"]
        Map["MapLibre GL Canvas"]
        Filter["Category Pills"]
        Dialogs["Handover Chat & Modals"]
        LiveHook["useLiveUpdates() Hook"]
    end

    subgraph Backend["API & Server Actions"]
        Resonance["Resonance Engine (math.ts)"]
        ChatAPI["/api/reunions/[id]/messages"]
        ProofAPI["/api/reunions/[id]/confirm-proof"]
    end

    subgraph SanityLake["Sanity.io Content Lake (ezgec0rq)"]
        Docs1["lostFoundItem"]
        Docs2["match"]
        Docs3["reunion"]
        Docs4["category"]
        Listener["Sanity listen() API"]
    end

    Client --> Backend
    Backend --> SanityLake
    SanityLake -->|Real-time stream| Listener
    Listener --> LiveHook
    LiveHook --> Map

The Anonymous Handover Chat

One of the key lessons during development was that strangers don't want to sign up or give out WhatsApp numbers to coordinate returning lost items.

LostNet introduces ephemeral, anonymous chat sessions bound to match pairs:

• Instant Meeting Chips: Quick-send preset chips ("Can we meet near 100 Feet Road?", "I have the item safely with me", "Left at front desk").
• Mutual Safety: Zero personal contact info required.
• Proof-of-Return Verification: Once returned, either party can mark the handover complete, triggering the reunion record on Sanity.

---

Sanity Project Details

• Sanity Project ID: ezgec0rq
• Dataset: production

LostNet treats Sanity not just as a static CMS, but as a live, structured civic ledger. The schema is organized into 5 primary document types:

src/sanity/schema.ts
├── category         # Taxonomic tags, emoji representations, and search aliases
├── lostFoundItem    # Geopoints, occurredAt, colors, materials, drop-off handover notes, status lifecycle
├── match            # Multi-document reference (itemA + itemB), score breakdown (category, geo, time, text)
├── reunion          # Confirmed returns, public celebration story, anonymous chat transcripts
└── settings         # Dynamic neighborhood radius (km), time window tolerance, confidence thresholds

Real-Time Live Sync with Sanity listen()

When an item is reported or resolved, other open screens reflect the update in real-time without refreshing:

// src/hooks/use-live-updates.ts
const subscription = client
  .listen('*[_type in ["lostFoundItem", "match", "reunion"]]')
  .subscribe((update) => {
    // Debounced callback to refresh map pins and resonance vectors
    onUpdate();
  });

Clean Provider Abstraction

All data operations sit behind a unified interface (src/lib/data/provider.ts). Setting DATA_PROVIDER=sanity in the environment points mutations and GROQ projections directly to Sanity's Content Lake:

*[_type == "lostFoundItem" && status != "resolved"] {
  _id,
  kind,
  title,
  description,
  placeLabel,
  "location": { "lat": location.lat, "lng": location.lng },
  occurredAt,
  colors,
  materials,
  status,
  imageUrl,
  handoverNote,
  "category": category->title
}

---

My Build Process

The AI-Native Pair Programming Loop

I built LostNet using Google Antigravity IDE, working in a continuous pair-programming loop with its agentic coding engine.

The Great Workflow Pivot: From Crypto-Nonsense to Human Reality

The most memorable moment during the build was when the code was technically "flawless", but the product was unusable in real life.

Initially, an over-engineered cryptographic handover system was generated:

• Owners had to set a secret question ("What color is the lanyard?").
• The server generated random salts and stored PBKDF2-SHA256 hashes.
• Finders had to input the claimant's answer into a simulated "Custody Desk" modal.
• The system generated QR code "Return Passes" styled like airline boarding passes.

It was impressive engineering, but when tested against common sense, a fundamental question arose: Who finding keys on a rainy sidewalk is going to quiz a stranger on cryptographic hashes or scan an airline boarding pass at a coffee counter?

A direct pivot prompt transformed the direction:

"don't just see the code problem, I think there is a workflow problem which doesn't happen in real life. The main part is handover — keep it very simple: kept in police station, or decided at place. Think of a way normal humans will actually use this."

That single reality-check reshaped the entire architecture:

• ✂️ Removed: Custody desk gatekeeping, PBKDF2 challenge hashes, and digital boarding passes.
• 🛠️ Introduced: 1-tap real-world drop-off categories (Dropped at known desk/police station, Holding safe, Left at spot).
• 💬 Added: Lightweight, anonymous in-browser chat with 1-click meeting coordination chips.
• 📍 Streamlined: Single-tap GPS capture and instant map pin placement.

---

Agent Session & Audit Logs

The complete multi-turn development log, automated testing passes, and workflow audits are preserved in the repository:

👉 View LostNet Human Workflow Audit Log on GitHub

---

Tech Stack

• Framework: Next.js 16 (React 19, Server Components & Route Handlers)
• Content Engine: Sanity Studio v6 & @sanity/client
• Geospatial & Vector Tiles: MapLibre GL
• Motion & Physics: Motion (Framer Motion v12) & Canvas Confetti
• Styling: Tailwind CSS v4
• Testing: Vitest (24/24 unit & resonance engine tests passing)