404 Food

Atlanta contributes to 48% of Georgia's food waste. Using AI, route leftover food to shelters based off needs and logistics, instantly.

Comment

Inspiration

Atlanta sends 50,000 tons of edible food to landfills every year while 820,000 residents depend on food banks, and demand has surged 40% in just two years. The FDA's 2-hour rule means perishable food left at room temperature beyond two hours must be discarded. That's all the time you have to get surplus from a venue to a shelter. We looked at where Atlanta's largest surplus actually comes from: stadium and arena event catering at Mercedes-Benz Stadium, Truist Park, and State Farm Arena; film and TV production catering across Georgia's 273 annual productions; and college dining halls (especially HBCUs), sitting directly adjacent to some of Atlanta's most food-insecure neighborhoods in Vine City and English Avenue. Every existing platform fights over restaurant leftovers. Nobody's building AI prediction around the event-driven, high-volume signals that make Atlanta unique, and nobody's racing that 2-hour clock.

What it does

AI agents negotiate in real time to route surplus food from venues to shelters before it spoils. A Donor Agent evaluates perishability, a Food Bank Agent checks shelter needs, and a Driver Agent optimizes routes using live traffic APIs. They reach consensus in under 30 seconds. Drivers scan a QR code to claim a unique delivery, each scanner gets a different route. Payouts are funded through corporate ESG budgets: $5 base + $0.20/lb + $1.50/mile.

How we built it

Frontend: React + Tailwind CSS + Framer Motion for a warm, editorial-style UI with real-time animations. Mapbox GL for live route visualization. Backend: FastAPI (Python) with async endpoints. Three OpenAI agents orchestrated in a pipeline, surplus prediction, shelter matching, and a multi-agent negotiation system that streams tokens in real time. Geo & Routing: Google Maps Geocoding API and Distance Matrix API for live traffic-aware travel times, with Nominatim (OpenStreetMap) as a fallback. Mapbox Directions API for turn-by-turn driver navigation. Driver Dispatch Queue: A backend session-based dispatch system where dispatched routes are registered via /api/dispatch, and individual drivers claim unclaimed routes via /api/claim-route. Exposed through ngrok so any phone on any network can scan the QR code and receive a unique delivery assignment. Data: Historical catering CSV data for stadiums, film sets, and HBCU dining halls used as context for AI predictions.

Challenges we ran into

Lots of challenges getting tooling to work with AI agents. Since API handling is exposed to each agent as a tool, its kinda hard to get each one to call the tools its supposed to while also being FAST. Luckily, GPT5.4 is pretty good at tool calls. Another challenge was the QR code mobile integration. It required orchestrating an event queue in the backend to ensure every "driver" got a unique route.

Accomplishments that we're proud of

Real multi-agent negotiation: Three GPT 5.4 agents genuinely argue with each other in real time, the Food Bank Agent pushes back on the Donor Agent's priorities, the Driver Agent factors in live traffic. This isn't a single prompt; it's a multi-turn debate that streams to the UI token by token. The QR dispatch system: Multiple drivers scan one QR code and each gets a different route. It's a simple idea but the implementation, session-based claiming, cross-device networking via ngrok, unique driver IDs, makes it feel like a real product. Live traffic integration: Google Maps Distance Matrix with departure_time=now gives real Atlanta traffic conditions. The routing matrix shows actual drive times, not estimates. Driver payout model that works: The ESG-funded formula ($5 + $0.20/lb + $1.50/mi) means a driver doing 4 deliveries a day earns ~$75. That's competitive with DoorDash without relying on charity.

What we learned

Multi-agent AI systems are powerful but hard to orchestrate getting three agents to genuinely negotiate rather than just take turns monologuing required careful context engineering and structured output schemas. The 2-hour rule makes this a real-time logistics problem, not a database problem. Every architectural decision made was driven by that clock. Atlanta's food surplus problem is hyper-local. The HBCU dining halls in the West End sit blocks from some of the city's most food-insecure neighborhoods. Routing that surplus to those communities isn't just efficient, it's deeply meaningful to Atlanta's history and culture.

What's next for 404 Food

  1. BeltLine delivery layer: Build a custom trail-segment graph for Atlanta's BeltLine so the routing agent can choose between a 6-minute car route (0.8 kg CO2) and a 12-minute e-bike trail route (0 kg CO2), and decide based on how much time remains in the pickup window.
  2. Partnering with restaurants and stadiums to make this work!

Built With

Share this project:

Updates