KAPHIY
AI chat ordering for a coffee shop, in production


AI chat ordering at Praliné Coffee House, with a real-time kitchen dashboard and sales recorded in the accounting ERP.
- Year
- 2026
- Context
- Praliné Coffee House, Baños de Agua Santa (Ecuador)
- Team
- 4 people
- Code
- GitHub
- automated tests
- 165
- releases with SemVer
- 11
- integrations: AI, real time and ERP
- 3
- kitchen board accessibility
- WCAG 2.1 AA
The system
Customers scan the QR code on their table and order from a PWA by chatting with an agent built on Gemini, which interprets natural language and assembles the order. The kitchen receives it instantly on a real-time board over WebSocket, and every sale is recorded in the client's accounting ERP so nothing is keyed in twice.
Architecture
Technical decisions
GitHub Actions runs 165 tests before changes are merged. Module contracts and the CI/CD pipeline keep production deployments repeatable.
What it does
- Customers scan the QR code on their table and order by chat, typing or by voice.
- An agent on Gemini 2.0 Flash, orchestrated with n8n, interprets the order and builds it against the real menu.
- The kitchen sees each order instantly on a board (KDS) connected over Socket.IO, with its status.
- Every sale is recorded in the client's accounting ERP (Contífico) without being keyed in again.
- Installable PWA with Serwist and an accessible kitchen board (WCAG 2.1 AA), in strict TypeScript with zero errors.
- 165 tests: 92 on the backend (Jest and Supertest), 28 on the PWA (Vitest and MSW) and 45 on the board (Vitest, Playwright and axe).
- 11 releases with SemVer, from v0.1.0 to v1.1.0, deployed on Vercel and Neon.
One concrete problem
Orders arriving five hours off
Order times showed up five hours off on the kitchen board: the columns were TIMESTAMP without a time zone and the server wasn't on Ecuador time. They were migrated to TIMESTAMPTZ with a versioned script (002_fix_timestamps_tz.sql): the database stores the absolute instant and each screen shows it in its local time.
Screenshots
The customer PWA: Praliné's menu and the way into the chat with the agent. The kitchen board (KDS): orders by table, in real time. The kitchen board's test suite, all green.
My part
Backend: data modelling with Prisma, the API and ordering logic, the real-time channel and the integration with the conversational agent. In a team of four (ARKASYS), over 14 weeks.
Stack
- NestJS
- Prisma
- PostgreSQL
- Next.js
- Socket.IO
- Gemini API
- n8n
- Docker
- GitHub Actions
