I added barcode scanning to my AI pricing app – 92% cheaper per scan

I built Underpriced AI, an app that identifies items from photos using Claude Vision and prices them using real sold data from 6 marketplaces (eBay, Poshmark, Mercari, etc.).

This week I added barcode scanning as an alternative input. The interesting part is the cost savings:

Photo scan: Claude Sonnet 4.6 Vision (~$0.05) + Perplexity Sonar (~$0.004) = ~$0.054/scan Barcode scan: UPCitemdb lookup (free) + Perplexity Sonar (~$0.004) = ~$0.004/scan For items with UPC/EAN barcodes (retail arbitrage, new-in-box products), this skips the vision model entirely. The UPC database returns the exact product, so confidence is actually higher than AI identification.

Technical details:

Browser-based barcode decoding via html5-qrcode (reads from a captured photo, not live video — getUserMedia is unreliable in Capacitor WebViews) UPCitemdb free tier (100 lookups/day, no API key) Same Perplexity + eBay Browse API pipeline for market research bwip-js generates barcode images for scan history when the UPC database doesn't return a product photo The app also runs on iOS and Android via Capacitor — same Next.js codebase. Camera access uses <input type="file" capture="environment"> which goes through the native OS camera picker, avoiding the getUserMedia permission nightmare on mobile.

Stack: Next.js 16 (App Router), TypeScript, Prisma + Neon PostgreSQL, Vercel, Claude Sonnet 4.6, Perplexity Sonar, Capacitor v8.

Other things shipped this week: investor forecasts dashboard with company valuations, 12 automated email sequences, operating cost tracking, MRR fixes.