About Verify AI

A trust layer for civic AI.

Verify AI is not just a chatbot. It is a browser-based civic AI safety system built to protect community users from harmful or misleading AI-generated information.

Mission

Why this exists

AI systems often generate confident-sounding responses that contain fake NGO phone numbers, incorrect hospital details, fabricated government schemes, or dangerously wrong civic guidance. In real community contexts, these errors can cause genuine harm.

Verify AI solves this by introducing a mandatory verification layer between AI generation and user delivery. No response reaches a user without passing an independent safety audit.

Architecture

The dual-agent pipeline

Verify AI separates the generation and verification responsibilities into two fully independent agents. The agent that creates a response is never the agent that approves it.

Generator Agent (Agent 1)

Produces a helpful, informative response to the user's civic question. Focuses on clarity and community guidance. Does NOT perform safety judgment.

Auditor Agent (Agent 2)

Independently verifies the generator's output. Checks for hallucinated contacts, unverifiable local info, overconfident claims, and misleading civic guidance. Returns structured PASS or BLOCK JSON.

Decision Engine

Enforces the auditor's structured output. Invalid or unparseable audit results default to BLOCK. No bypass possible.

Safety Model

PASS or BLOCK

The auditor returns a strict structured JSON object. The decision engine acts on it without interpretation.

PASS

General civic guidance
Safe educational content
No specific unverifiable details
Low-risk informational answers

BLOCK

Hallucinated phone numbers
Fake NGO or clinic contacts
Unverifiable local addresses
Overconfident civic claims

Default rule: When the auditor is uncertain, it blocks. System safety takes priority over completeness.

Technology

How it's built

Frontend

Next.js App Router, TypeScript, Tailwind CSS

UI Layer

shadcn/ui, Framer Motion, GSAP

Backend

Python FastAPI, Pydantic, Uvicorn

AI Engine

Google Gemini API (dual-agent)

Validation

Strict JSON schema, BLOCK-first fallback

Security

API keys server-side only, no login required

Future Scope

What comes next

→Google Search grounding for real-time civic fact verification
→Google Maps integration for location-based community queries
→Multi-auditor system (Fact, Bias, Logic auditors)
→Confidence scoring and audit transparency logs
→LiveKit realtime voice interaction
→External civic databases for ground-truth verification