Pramaan
Pramaan verifies whether a claim about the physical world actually matches reality — the missing layer where fraud hides in lending, insurance, and asset finance.
Core trick
Compares a live camera feed against a stated claim in Gemini Live, flags mismatches in real time, and falls back to on-device Gemma when offline.
Description
Pramaan verifies whether a claim about the physical world actually matches reality — the missing layer where fraud hides in lending, insurance, and asset finance. Online (Gemini Live API), the agent watches, listens, translates Hindi↔English, and proactively flags discrepancies in real time. Offline, the same agent runs on-device on Gemma — counting what's actually present against the claim, recovering from bad frames, and deferring to a human. Real-time when online, on-device when offline. Demonstrated on rural field lending; the same engine extends to insurance, subsidy, and inventory checks. Built on the Real-Time Multimodal Interaction track; also competing for Best Use of Gemma 4.
README excerpt
# Pramaan — proof that the work actually happened **A verification layer for the physical world** — does a claim match what the camera actually sees? Real-time or on-device. **Demonstrated on field lending** (our wedge; the same engine extends to insurance, subsidy & inventory verification). **Track:** Real-Time Multimodal Interaction (Gemini Live API) · **Also competing:** Best Use of Gemma 4 (Local-First Agents on-device) **Live demo video:** _<paste link>_ · **Run:** `npm start` → http://localhost:8790 --- ## The problem AI agents can act — approve, file, notify — but **none can confirm the physical world matches what was claimed.** That gap is where fraud hides: across lending, insurance, subsidies and asset finance, decisions rest on a field photo nobody can trust — ghost businesses, recycled photos, phantom collateral. We prove it on **rural field lending** — India's highest-fraud, lowest-connectivity, highest-impact wedge. ## What it does — and why it's different An agent that verifies a claim against what the camera **actually sees**, and acts on its own: - **Proactive.** It isn't asked — it *interrupts* the moment reality contradicts the claim. - **Real-time & multimodal.** Sees + hears + translates (Hindi ↔ English) live. It does **not** work typed into a chatbox — the interaction *is* the product. - **Works offline, on-device.** No signal? The same agent runs on the phone — counts assets, recovers from bad frames, and **defers to a human** instead of guessing.