MealFlow

MealFlow helps organizations cut food waste by collecting employee meal preferences, using AI insights, and securely verifying meal collection with QR passes.

Comment

Inspiration

Office and factory canteens routinely over-prepare food because they have no reliable signal of who's actually eating the next day — leading to daily waste, and sometimes running short for employees who do want to eat. We wanted a simple way for employees to signal their meal plans a day ahead, in their own words, and have that translate directly into what the kitchen prepares. ...as we usually write... This entire build was also an experiment in AI-native development — GPT-5.6 shaped the system design before a single line was written, and Codex then implemented it feature by feature.

MealFlow

AI-powered meal planning — architected with GPT-5.6, built with Codex.

Built with Codex Planned with GPT--5.6

What it does

MealFlow is a role-based web app for company canteens:

  • Employees get asked every evening what they want for tomorrow's breakfast/lunch/snacks. They can pick from the menu, skip a meal, or just type a reason in plain language — "I'll be on leave tomorrow" — and the AI converts that into a structured skip/preference automatically.
  • Master Chefs publish the next day's menu with item-level capacity limits (e.g. only 100 biryanis), so the kitchen never over- or under-cooks.
  • Main Admins onboard the company, import employees via CSV/XLSX, set free/paid/subsidized meal pricing, and see participation + waste analytics.
  • Chef Helpers verify meal pickup via a signed, single-use QR pass — no personal data stored in the code itself.
  • Head Admin monitors all companies on the platform and handles account recovery, without ever touching a company's day-to-day meal data.

Every confirmed selection generates a QR meal pass; scanning it at pickup marks the meal collected exactly once, preventing double-serving.

How we built it

  • GPT-5.6 for architecture and product planning — designing the normalized database schema, mapping out the five-role workflow (Head Admin, Main Admin, Master Chef, Chef Helper, Employee), and working through edge cases like atomic capacity reservation and AI confidence-tiering before any code was written
  • Codex as the primary coding agent — implemented nearly every feature end-to-end, from auth and admin dashboards to CSV/XLSX employee import, menu publishing, and QR meal-pass generation, built one scoped feature at a time and verified against the live app and database
  • Next.js + TypeScript + Tailwind CSS for the responsive web app
  • Supabase for authentication (passwordless email OTP), Postgres database, and row-level security so each company only sees its own data
  • Resend for invite and reminder emails
  • Gemini API for natural-language meal-preference understanding, using a confidence-tiered approach: high-confidence requests apply automatically, medium-confidence ones ask for one-tap confirmation, and unclear ones prompt a clarifying question
  • Deployed on Vercel

We designed the database around normalization and audit logging from the start — every AI decision, admin edit, and QR scan is logged with the original input preserved, so nothing is silently overwritten.

AI-assisted development

This project leaned heavily on two AI tools, each for a distinct role:

GPT-5.6 — architecture and planning Used before any code was written to design the normalized database schema, define the five-role workflow (Head Admin, Main Admin, Master Chef, Chef Helper, Employee), reason through tricky edge cases like atomic capacity reservation to prevent overselling, and design the confidence-tiered logic for natural-language meal preferences.

Codex — implementation Used as the hands-on coding agent for nearly the entire build — auth, CSV/XLSX import, menu publishing, QR generation, Supabase migrations — one scoped feature per prompt, each verified against the running app and database before moving forward.

Every AI-generated change was reviewed and tested manually; AI accelerated implementation, but design decisions and verification stayed human-driven.

Challenges we ran into

  • Preventing overselling: worked through this with GPT-5.6 to design an atomic database update instead of a check-then-write pattern, then had Codex implement it directly.

  • Free-tier email limits: both Supabase's built-in mailer and Resend's default sending address restrict who you can email during testing, which we worked around with a custom SMTP domain and a safe, clearly labeled demo-login mode for reviewers.

  • Keeping AI answers safe and scoped: making sure the natural-language assistant only ever acts on meal preferences — never guesses silently on low-confidence input, and never touches another employee's data.

  • Data isolation: making sure one company's data is never visible to another, enforced at the database level with row-level security rather than just in the UI.

What we learned

Building this reinforced that the "boring" plumbing — atomic transactions, audit logs, role-based access, rate limits — matters more for a real operational tool than the AI feature itself. We deliberately built the plain forms-and-buttons version of every flow first, and only layered AI on top once the underlying data flow was solid.

Acknowledgments

Built with the assistance of GPT-5.6 (architecture & planning) and Codex (implementation).

What's next

  • Push notifications and WhatsApp reminders alongside email
  • Demand forecasting using historical selection data
  • Native mobile app (React Native/Expo) sharing the same backend and rules
  • Support for factories, schools, and event receptions using the same configurable data model

Built With

  • ai-prediction
  • authentication
  • codex
  • csv
  • food-waste-reduction
  • gemini
  • gpt5.6
  • indexing
  • javascript
  • jwt
  • next.js
  • qr-code
  • rag
  • rbac
  • react
  • resend
  • responsive-web-design
  • rest-api
  • rls
  • smtp
  • supabase
  • supabaseauth
  • typescript
  • vercel
  • xlsx
Share this project:

Updates