Treat content as data
Product descriptions are parsed into explicit attributes rather than stored as undifferentiated marketing copy. That makes discovery, comparison and programmatic pages possible.
Indian
A solo-built discovery platform that turns fragmented Indian specialty-coffee catalogues into structured, searchable data.
product · live
Founder, product owner and sole builder · Updated Aug 2026
- Data systems
- AI extraction
- SEO
- Product
01 · Problem
India's specialty coffee is spread across dozens of independent roaster sites with no shared catalogue or consistent product data.
02 · System
Scrapers collect source catalogues, language models normalize inconsistent descriptions into a common schema, and automated QA and publishing loops keep the catalogue useful.
03 · Outcome
92 roasters, 1,400+ SKUs, 2,000+ monthly uniques and 1,000–3,000 weekly active users through organic discovery.
Context
Coffee buyers often know a flavour, process or origin they want, but not which roaster currently carries it. Roaster sites describe similar attributes in different ways, making comparison difficult.
The product had to work as both a consumer discovery experience and a continuously maintained data system. Manual catalogue entry would have made the project impossible to operate alone.
Architecture
- 01Source discovery and scheduled catalogue scraping
- 02LLM-assisted extraction into a shared product schema
- 03Normalization and validation for origins, processes and tasting notes
- 04Supabase as the structured content and application data layer
- 05Next.js discovery pages designed for search and browsing
- 06Automated refresh and QA workflows orchestrated with n8n
Decisions & trade-offs
Keep humans at the exception layer
Automation handles repeatable extraction and refresh work. Review effort is reserved for ambiguous records and schema failures instead of every item.
Build for organic utility
Pages are organized around questions coffee buyers already ask, so the same structured data serves the interface and the search surface.
Evidence
- 92 roasters represented in one discovery layer
- More than 1,400 coffee SKUs normalized
- More than 2,000 monthly unique visitors
- 1,000–3,000 weekly active users, acquired organically
What this taught me
- The difficult part of an AI extraction system is not the prompt; it is the schema, the failure taxonomy and the review loop.
- A useful taxonomy can become both product navigation and a scalable distribution system.
- Solo operation is viable when automation is designed around exceptions instead of pretending exceptions do not exist.
Follow the system
Connected work is resolved from the shared content graph.
Discusses projectI Built a Data Pipeline for 92 Coffee Roasters. I Couldn't Write a For-Loop Six Months Ago.WritingUsed inNext.jsToolkitUsed inSupabaseToolkitUsed inn8nToolkitCreated fromData-normalization field guideToolkitDiscusses projectI Almost Deleted 16 Specialty Coffees From My Database. Two Bugs Were Conspiring in Silence.WritingDiscusses projectI Went Through 5 Database Schemas Before I Wrote a Single Line of Frontend CodeWritingDiscusses projectTailwind 4 Gave Me Design Tokens and Emotional DamageWritingDiscusses projectAfter Rebuilding My Code 3 Times, I Finally Finished My Damn Project (Well, Half of It)WritingDiscusses projectI Almost Didn't Launch IndianCoffeeBeans.com — Until ChatGPT Bullied Me Into ItWritingDiscusses projectI Built IndianCoffeeBeans.com Using ChatGPT, Claude, and Caffeine at 4AMWritingDiscusses projectA Single `null` Killed Every SEO Tag on My Site. The Fix Was Three Characters.Writing