Badger

What it does

Your landlord won't return your deposit. You Google it and get a wall of legal jargon, or worse — an AI chatbot that confidently cites a law that doesn't exist.

Badger gives everyday people verified legal answers in plain language. Pick your country, describe your situation, and get structured guidance: your rights, key deadlines, concrete next steps, and real legal citations you can click to verify yourself. If you need to send a formal letter, the built-in editor walks you through it step by step and exports a PDF — all without ever sending your personal data to a server.

We cover 6 countries (UK, US, Taiwan, Japan, South Korea, Singapore) across 4 common scenarios: rental deposits, consumer refunds, unfair dismissal, and data breaches.

How we built it

The core innovation is the Citation Index — a curated, human-verified index of real laws per country. Instead of RAG or vector search (which still hallucinate), the AI model is constrained to only cite references present in the index. This eliminates the #1 trust problem with legal AI: made-up references.

The stack is Next.js 16 + TypeScript + Tailwind, powered by Google Gemini 3.1 Flash Lite. The interactive 3D globe (Three.js + React Three Fiber) makes country selection feel tangible. GSAP handles all transitions. The letter editor is a Typeform-style guided flow that runs entirely in the browser — personal details never leave the user's device.

Novus analytics tracks the full user journey: onboarding, country/topic selection, query submission, workflow completion, and user feedback signals. We used Novus insights to identify and fix rage click patterns on our loading states and dead clicks on the letter editor.

What we learned

• Citation indexes beat RAG for trust-critical domains. When wrong information has real-world consequences, constraining the model's reference pool is more reliable than hoping retrieval finds the right chunk.
• Loading UX matters more than loading speed. Our API responds in 5-10 seconds, but without clear progress indication, users rage-clicked the button at a 74% frustration rate. Adding estimated wait time and visual disabled state cut this dramatically.
• Privacy messaging must be precise. "Zero personal data" sounded good but was technically inaccurate — the letter editor does handle personal data, just locally. We learned to say what we actually do: "Your data stays local."

Built With

• google-gemini
• gsap
• jspdf
• next.js
• novus
• react-three-fiber
• tailwind-css
• three.js
• typescript