ThriftLens

One-liner: Snap a photo of a thrift find and get a HOT/GOOD/PASS deal score in under 20 seconds — Claude Vision identifies the item, live eBay comps price it, and the app tells me whether the margin is real before I ever put it in the cart.

Origin (the Build Loop)

Because I was reselling thrift finds, I kept hitting the moment in the aisle where I had to guess at an item’s resale value and margin from memory, so I built a mobile PWA that photographs the item, identifies it with Claude Vision, pulls live eBay comps, and scores the deal, that turns a gut guess into a market value range and a profit estimate on the spot. It taught me how to keep an AI pipeline cheap and fast enough to actually use standing in a store — vision once, text-only after, and cache the expensive external call.

The problem

Sourcing thrift items to flip means making a buy/pass call in seconds, in the aisle, on your phone. The information I actually needed — what is this, what does it sell for right now, and what’s left after fees and shipping — lived across three tabs and a lot of guesswork. I wanted one motion: point the camera, get a verdict, decide.

What it does

• Camera or library upload with preview, retake, and “choose different”
• Claude Vision identification: brand, model, condition, era, category
• Live eBay Browse API comps (active listings), cached 24h to stay fast and cheap
• Market value range (low/high) plus a net profit estimate after platform fees and shipping
• Deal score — HOT / GOOD / PASS — with a confidence note
• Selling tips: best platforms, listing keywords, authenticity warnings
• Per-session scan history with client-side filters and a “scan again” CTA
• Installable PWA (iPhone Safari, Add to Home Screen)
• No account required; images are never stored

How it’s built

• Stack: Next.js 16 (App Router), Claude Sonnet 4.6 (Vision + Text) via @anthropic-ai/sdk, eBay Browse API, NeonDB (serverless PostgreSQL), Tailwind CSS v4, native service worker + manifest for PWA.
• Notable engineering:
  • Two-call AI pipeline that only pays for vision once. The photo hits Claude Sonnet 4.6 Vision one time to produce a structured IdentificationResult (including an ebay_search_query). The final analysis — deal score, market value, profit — is a text-only Sonnet call that reuses the identification and comps without re-sending the image, roughly 10x cheaper than a second vision call.
  • Comps cache keyed by SHA-256 of the search query. Each identification’s eBay query is hashed; a comps_cache lookup in NeonDB serves a hit or falls through to the eBay Browse API on miss, then caches the result for 24 hours. The expensive external call happens once per distinct item, not once per scan.
  • eBay OAuth client-credentials flow with a cached token, wrapped in a small Browse API client.
  • eBay API proxy client (Sept 2026) — extraction of the eBay integration into a dedicated proxy layer, consistent with the shared-endpoint pattern used across the eBay-powered trackers.
  • Profit & loss engine turns raw comps into a net estimate after platform fees and shipping — the number that actually decides the buy.
  • Honest data labeling: comps are active listings (asking prices), not sold prices, and the UI says so plus surfaces what the confidence figure means.
• Architecture: POST /api/scan runs the pipeline end to end — Vision identify → comps cache lookup → eBay Browse on miss (cache 24h) → text-only analysis → NeonDB insert → redirect to /results/[id]. Sessions are a UUID in localStorage; there’s no auth in beta. A /api/migrate endpoint initializes the schema.

Proof points

• Verdict in under 20 seconds from photo to HOT/GOOD/PASS.
• Two model calls per scan, only one of them vision — the text-only analysis call is ~10x cheaper than re-sending the image.
• eBay comps cached 24h per item via SHA-256 query hash; eBay’s free tier is 5,000 calls/day, so caching keeps the app well inside it.
• Zero image storage and zero accounts — privacy is a stated product feature, not an afterthought.

What to show

• Demo: Live PWA (private repo, public demo when deployed). Best shown on an actual phone — open in iPhone Safari, Add to Home Screen, scan a real item.
• Visuals needed: phone screen recording of the scan → processing stepper (Identify → Comps → Analysis) → results card with the deal badge; a diagram of the /api/scan pipeline; a results screenshot showing the profit estimate. Blur any real API keys if the network tab is visible.

Cross-links

• Same photograph-it-and-extract Vision pattern as the DJ paperwork PWA, COMPLETE-PAPERWORK — point the camera at a real-world artifact, let Claude Vision turn it into structured data, act on the result. ThriftLens applies that pattern to resale sourcing instead of wedding paperwork.
• Shares the “AI says: is this actually a deal?” shape with public-storage-bot (rates → Claude verdict) and POLYBOT (market price vs. AI estimate → trade/skip).

Case-study angle

Alex built a real-time, cost-aware AI product — vision once, text after, external comps cached — that survives being used one-handed in a thrift-store aisle, and turned “I think I can flip this” into a defensible profit number.