AeroBites

TL;DR

• Where the drone-delivery idea started: at SCU ACM’s AWS x INRIX Hack in October 2025, my five-person team set out to let UC Santa Cruz students order dining-hall food by drone.
• I authored 13 of the team repo’s 18 commits, including every commit that changed code: the app, the menus across 20 dining locations, order tracking, the delivery map and the chat assistant.
• We had no drone, so I flew one in simulation. Three months later I rebuilt the concept on live data as SlugBites, and our team won two awards with it.

The problem

Campus dining is a walk away from where students actually are, and a sponsor event built around AWS and mobility data was a good excuse to ask what drone delivery would take: an ordering app, an agent that can take an order in plain language, and a way to show a flight.

What I built

• App. React 18 with Tailwind and TanStack Query: menus with dietary filters, a cart with a dorm selector, order history, an order tracker and a delivery map.
• Back end as code. An AWS architecture written in Serverless Framework: four Lambda HTTP functions (an AI agent, recommendations, natural-language ordering and a Bedrock agent), a DynamoDB table for agent memory keyed by hash and range, and an S3 knowledge-base bucket.
• Agent. A Bedrock function that wraps Claude 3 Sonnet in an ordering persona.
• Simulation. On the final morning I stood up Microsoft AirSim’s prebuilt Unreal Engine environments, flew a quadrotor by hand with RGB, depth and segmentation views, then launched AirSim’s Neighborhood map and started a screen recording.

Key decisions

• Decision: infrastructure as code for the whole back end. Why: it is reviewable and reproducible from one file. Trade-off: it stayed undeployed at the event; the IAM setup was still outstanding.
• Decision: a mock data layer under the UI. Why: the app demos end to end without a deployed back end. Trade-off: the demo ran on mock menus, which is exactly what SlugBites fixed.
• Decision: show the drone in a simulator. Why: no hardware, and a real flight model beats an animation. Trade-off: manual flights, not autonomous delivery.

The hard part

Demoing drone delivery with no drone. The team had no hardware. In the last hours I set up AirSim’s Unreal Engine environments on my own machine and flew a quadrotor with multi-camera sensor views, then switched to the Neighborhood map and started a screen recording 46 minutes before the team’s final commit.

What I learned

AeroBites’ back end stayed undeployed and the UI ran on mock data. Both became the brief for round two: SlugBites scraped real menus and ran its own API from day one, and six weeks after that I was flying scripted missions from a web console in Riptide.

What I’d do next

This version is archived; the idea continued in SlugBites and Riptide.

Links

• Code: github.com/versean/AeroBites (team repo)
• Devpost: AeroBites
• Round two: SlugBites · Round three: Riptide