Case study · AI pipeline
Email-to-Wardrobe Pipeline
AI pipeline that pulls 13 years of fashion purchases from Gmail into a digital wardrobe. 1,191 emails · $1.29 total · 102 items.
Role
Solo builder (product + engineering)
Timeline
1 week, research through live deploy, Mar 2026
Stack
Next.js 15, TypeScript, Supabase, Claude Haiku, Gmail API, Zod, Vercel
Scope
4 personas, 6-phase plan, brand, 7-layer extraction, cost analysis, Google OAuth
Why this exists
People own more clothes than ever and can’t find half of them
During the war, I was stuck at home, reorganizing my closet, and I had this unsettling realization: I didn't know what I owned. What I paid, when I bought things, whether I'd actually worn them.
53 items purchased per year on average, 4× more than in 2000. 40% never worn. ~100 hours a year in "wardrobe panic."
Scraping fashion receipts from email now covers far more than a few years ago, including online and in-store digital receipts.
Wardrobe apps (Whering, Acloset, Cladwell, Alta…) need you to photograph every item before you get value. Almost nobody finishes.
Receipts already lived in email, and after COVID even physical stores push emailed receipts.
I wanted to know if that signal alone could build the closet without the photo grind.
So I built Kasane.
The product
Connect Gmail, wait ~3 minutes, see your wardrobe
Sign in with Google → inbox search → filter noise → purchase emails to Claude → structured save. You get brand, category, size, color, prices, dates, images.
Four currencies stored as-is; display converts with cached FX (e.g. ILS totals across USD / GBP purchases).
How it works
Pipeline architecture
1,191 emails in → 102 items out → $1.29 total. Claude is the expensive step, so the pipeline treats the LLM as a last resort, not a first pass.
Gmail search
6 parallel queries: purchases, Shopify, Global-e, phrases, Hebrew, receipts.
Subject pre-filter
Cheap heuristics before any model call.
Claude extraction
Structured JSON + Zod; retries and per-email cost tracking.