Cocina OS

Inspiration

Last summer I went to Mexico City and lived with my aunt for 6 weeks. She runs Aventura Gourmet, a chain of 3 restaurants (Motín Juárez, Motín Roma, Quesería). I went because I love cooking and wanted to work in her kitchens and connect with my Mexican roots. But I'm from Silicon Valley, so I can't help but notice tech stuff. And the back office was a mess. Paper invoices everywhere, costs tracked across like 20+ Google Sheets that barely made sense, managers spending hours every week typing numbers from paper into spreadsheets by hand. I looked into real solutions like MarketMan, Oracle, Bind ERP, but they were way too expensive for a 3-restaurant chain. I made a whole pitch deck and presented a modernization plan to Aventura Gourmet before left. But the costs & time requirement to set up killed it. Few months later, back in San Jose, I was like... I'll just build it myself.

What it does

You take a photo of a paper invoice or upload a PDF. AI extracts everything: supplier, date, line items, quantities, unit prices, taxes, all in seconds. You review it, fix anything that looks wrong, and hit submit. It saves to Google Sheets (because that's what they're used to) and to a real database that powers analytics dashboards. The dashboards let you see spending by supplier, track ingredient costs over time, and compare across all 3 restaurants. That visibility didn't exist before. The system also handles Mexican date formats, pesos, and 100+ weird unit variants that Mexican suppliers use (CHAROLA, MANOJO, etc.). It's built to deal with the inefficient and old way that business is done in Mexico.

How we built it

Next.js 15, React 19, TypeScript, Tailwind, Supabase for the database, OpenAI GPT-4.1-mini for extraction, Google Sheets API for compatibility, hosted on Vercel. I planned the architecture, picked the stack, found the APIs, and designed the workflow. Used Claude Code for a lot of the coding, same way any startup would use AI tools to ship faster. In the process of building this, I greatly improved the way I worked with these new types of AI tools.

Challenges we ran into

Mexican invoices are very unpredictable. Handwritten, blurry scans, inconsistent formats, abbreviations for units that don't exist anywhere standard. Building a unit normalizer that handles 100+ variants was painful. Getting AI to reliably extract data from photos of paper (not clean PDFs) took a lot of prompt engineering. Also, the users aren't technical at all. The whole system had to be in Spanish, simple, and forgiving. to solve this, I built an interactive tutorial, inline editing and auto-saving drafts so nothing gets lost.

Accomplishments that we're proud of

400+ invoices processed in the first month. 20+ hours of manual work saved per month. What took 15 minutes per invoice now takes under a minute. Batches of 100 invoices process in ~15 seconds because all the API calls run in parallel. My aunt loved it so much she paid for it. I have a paying customer at 17.

What we learned

The biggest decisions weren't technical, they were about understanding how people actually work. I watched managers, but also myself, process invoices by hand for weeks before writing any code. That's why it saves to Google Sheets and a database. You have to meet users where they are. Also learned firsthand how massive the tech gap is. 640,000+ restaurants in Mexico, most running their back office on paper. Affordable modern tools built for how they actually operate basically don't exist.

What's next for Cocina OS

Going back to Mexico City this summer. Adding POS integration for sales data, recipe costing, and sales-vs-cost analysis. Then making it multi-tenant so I can sell it to other restaurants. My aunt has connections in the industry and is helping me reach more businesses. Long term, bring affordable back-office software to the hundreds of thousands of restaurants across Mexico and Latin America that can't afford enterprise solutions.

Built With

• googlesheetsapi
• next.js
• openai
• react
• supabase
• typescript
• vercel