A food-ordering app with AI image search — point a camera at a dish and it finds it. A five-person OU research project that placed Top-3; I built the backend, the admin, and the mobile client.
YumYard is an online food-ordering app with a twist — alongside keyword search, you can point your camera at a dish and an AI model identifies it, then finds it on the menu.
It was a five-person university research project. I owned the engineering: an Express + TypeScript REST backend on MongoDB, a Next.js admin panel, and the Flutter client — three codebases wired to the ML team's recognition model through an image-search service.
The model, trained by the ML team on 30,612 images across 80 food categories, reached 92.10% accuracy — and the project placed Top-3 at OU's Student Scientific Research Competition 2024–2025, with a cash award of nearly USD 1,000.
I built the parts that move data and serve users — a REST backend, an admin panel, and a mobile client. The ML team built the recognition model; an image-search service is the seam between the two.
Firebase Auth (email / Facebook / Google) · keyword + image search · cart · order tracking
NextAuth JWT sessions · 3 management modules
REST API · JWT auth · food CRUD · order lifecycle · cart · favorites · reviews
30,612 images · 80 categories · 92.10% test accuracy
The model gets the attention, but a demo only ships if everything around it works. These four are the engineering I owned.
An Express + TypeScript backend on MongoDB Atlas, Swagger-documented — JWT auth with User and Admin roles, food CRUD, the full order lifecycle, cart, favorites, and reviews. Organized into User, Restaurant, Order, and Image-Search areas.
The image-search service proxies a photo from the Flutter client to the recognition model and maps the result back to a real menu item — with a confidence-aware fallback UI for the times the model isn't sure.
The Flutter client carries Firebase Auth across email, Facebook, and Google; search by keyword or by image from the camera or gallery; and cart, checkout, and order tracking. The image-search path renders results with a confidence-aware fallback.
A Next.js admin panel with NextAuth JWT sessions and three management modules — the operations side of the product, kept as its own codebase, separate from the customer app.
YumYard was a five-person OU research project. The 92.10% accuracy is the ML team's model result; the three codebases — backend, admin, and mobile — are my engineering contribution. Snapshot from the project, Jan–Feb 2025.
