Grail Hunter Lite - AI Vintage Fashion Authentication

Inspiration

The $50B+ resale market has a trust problem. Authenticating vintage clothing costs $50-200 per item and is gatekept by a handful of experts. Meanwhile, counterfeit vintage floods Grailed, Depop, and eBay — costing buyers billions annually. We asked: what if your phone camera could do what only trained authenticators can?

## What it does

Point your camera at any thrift store find. Grail Hunter runs a 4-phase forensic analysis:

1. Visual Thought Signature — Examines stitching patterns, hardware, fabric weave
2. Taxonomy Analysis — Identifies brand, era, category, rarity tier
3. Market Delta — Estimates current resale value from comparable sales
4. Authentication Verdict — Delivers confidence score with reasoning chain and red flags

The app includes a real RN/WPL Dating Engine — FTC Registration Numbers used by actual vintage dealers to date garments. This isn't AI hallucination; it's regulatory science: year = 1959 + floor((RN - 13670) / 2635).

## How we built it

5 Gemini APIs power distinct features across the app:

| API | Model | Feature | |-----|-------|---------| | Vision + Extended Thinking | gemini-3-flash-preview | Forensic scan with thinkingLevel: high + structured JSON output | | Search Grounding | gemini-3-flash-preview | Intel tab — real-time market Q&A backed by Google Search | | Maps Grounding | gemini-2.5-flash | Map tab — discovers nearby vintage/thrift stores | | Text-to-Speech | gemini-2.5-flash-preview-tts | Audio briefings of scan results (Kore voice) | | Veo 3.1 | veo-3.1-fast-generate-preview | Cinematic product reels for social sharing |

Stack: React 19 + TypeScript (strict mode), Vite, Tailwind CSS. Zero runtime dependencies beyond React and the Gemini SDK. 77 tests across 13 files. Works fully in simulation mode without an API key.

## Challenges we ran into

• Extended thinking latency — thinkingLevel: high produces incredible forensic reasoning but takes 15-45s. We built a multi-phase HUD animation that displays forensic phase labels to keep users engaged during the wait.
• Structured output schema design — Getting Gemini to reliably return typed JSON with confidence scores, red flag arrays, and material composition required careful schema engineering with responseMimeType: application/json.
• Maps grounding fallback — The Maps API needs browser geolocation. We built graceful degradation to SF coordinates when permission is denied or unavailable.
• XSS prevention — AI chat responses render as HTML for markdown formatting. We wrote a custom escapeHtml() sanitizer that runs before markdown-to-HTML conversion to prevent injection from model outputs.

## Accomplishments that we're proud of

• RN/WPL Dating Engine — Real textile forensics, not just an AI wrapper
• 5 distinct Gemini surfaces in one cohesive app — most entries use 1-2
• 77 passing tests with zero TypeScript errors in strict mode
• Simulation fallback — every API gracefully degrades when no key is set
• Badge gamification — 6 collectible achievements that create retention loops

## What we learned

• Extended thinking is a game-changer for structured reasoning — the forensic chains it produces are genuinely expert-level, not generic summaries.
• Search grounding transforms a chatbot into a research analyst — the Intel tab answers market questions with cited sources, not hallucinations.
• JSON schema + responseMimeType makes AI responses deterministic enough to drive real UI components like confidence rings and red flag cards.

## What's next

• Batch scanning — scan entire thrift store racks and get a ranked "hunt list"
• Price tracking — monitor items over time with search grounding for live comp data
• Community verification — let authenticated experts confirm or challenge AI verdicts
• Mobile app — native camera integration for faster capture

Built With

• gemini
• gemini3
• geminimapsgrounding
• geminisearchgrounding
• geminitts
• geminivision
• react
• tailwindcss
• typescript
• veo3.1
• vercel
• vite