TreasureLens AI

🏠

Inspiration

Helping clear out a grandparent's house, we kept hitting the same wall: "Is this worth anything?" We almost donated a lamp worth hundreds. The hard part of downsizing isn't selling — it's knowing what's valuable before it's gone. We wanted to put an AI estate-sale expert in everyone's pocket. 💡

What it does

From a single room photo, TreasureLens AI: Detects every item and estimates total room value ($4,870 across 9 items in our sample) Value X-Ray — darkens the room and overlays live price tags on every object Hidden Treasure detection — flags heirlooms to appraise before selling, so nothing is under-sold Liquidation Plan — routes each item to Sell / Bundle / Appraise / Donate / Recycle / Keep Auto-drafted listings with title, price, floor, and best-fit marketplace Negotiation copilot — accept / counter / decline against your floor price Human approval gate on every listing and offer For downsizers, estate executors, and anyone clearing a garage or attic. The question it answers isn't "how do I list this?" — it's "what is this whole room worth, and what should I do with it?" 🛠️

How we built it

We modeled the agent as a five-stage pipeline, each stage a swappable module behind one shared interface: Stages 4–5 run for real via a ResaleAgent interface (draftListing() + negotiate()) in src/lib/agent.ts, with a fully offline implementation. Stages 1–3 use prepared outputs so the demo is fast and reliable — a live vision model (GPT-4o, Gemini) and real sold-price data drop in with no UI changes. Valuations use an explainable model, not a black box: Each item ships with a value range and confidence score. Low-confidence items are gated to "Appraise First" instead of guessed at. Stack: React + Vite + TypeScript + Tailwind + Framer Motion on our "Electric Slate" design system, built with Claude Code as our AI pair-programmer. Built responsibly: approval-first, privacy-aware image handling, and honest ranges with confidence so heirlooms aren't under-sold. 🧗

Challenges we ran into

Making a scripted demo feel live — animation timing, the climbing value counter, and the X-Ray reveal all had to land in seconds Designing five swappable stages so a real model slots in without a rewrite Keeping valuations honest with ranges + confidence instead of fake-precise numbers 🏆

Accomplishments we're proud of

End-to-end flow that reads as a real agent: photo → detection → valuation → plan → listings → negotiation Working Value X-Ray and real negotiation math (accept / counter / decline) Judge-ready interface with a consistent design system 📚

What we learned

Decision-making was the hard part, not listing creation. The product is stronger as a resale advisor than a listing generator. The human-in-the-loop approval gate is a feature, not a limitation. Designing swappable stages early makes "sample data now, real model later" a real path, not a promise. 🚀

What's next

Live vision model + real sold-price data (eBay sold listings) behind the existing five stages Real marketplace cross-posting (Facebook, eBay, OfferUp, Poshmark) Buyer-message integration with approved negotiation rules and pickup coordination © 2026 Enrica Garrino. TreasureLens AI — original concept, design, and source code. All rights reserved.