GILGAMESH

Inspiration

Airport security is slow and annoying. We've all stood in that line while the belt stops because one bag needs a second look. Then we thought about the officer behind the screen, looking at X-rays for hours with a few seconds per bag. That's tiring, and tired people miss and slow things. So we built a tool that spots the risky bags for them, so the line moves faster and they can focus on the bags that matter.

What it does

Upload X-ray scans, and two AI models check each one for knives, blades and scissors.
Each scan becomes a bag on a 3D belt. Clear bags keep moving. Flagged bags get pushed aside: red for a likely threat, orange for "look closer."
Click a held bag to see a 3D view of the scan and the owner's info.
The officer makes the final call, and every decision is saved to History.

How we built it

AI: We trained a YOLO26s model on Roboflow and paired it with an open-source YOLOv8s model. We merge their results so one catches what the other misses.
In the browser: Both models run locally with onnxruntime-web, so scans never leave the computer.
3D: three.js for the belt and the scan view.
App: React + TypeScript, with a small Node/Express backend for History.

Challenges we ran into

Our dataset named five classes just "0" to "4," so we had to figure out what each one was.
The model isn't perfect yet. Its precision is around 80–90%, and it still misses some prohibited items.
We ran out of credits for training our models.
At first, the 3D belt and the Console didn't agree on which bags were flagged. We rebuilt it so both use the same result.
Our first UI was confusing, so we started over with a cleaner one.

Accomplishments that we're proud of

As an all-freshman team, we built a working product, not a mockup.
Trained two AI models running fully in the browser.
A 3D lane that makes the AI's decisions easy to see at a glance.
A good-looking and easy to understand UI/X

What we learned

Data matters more than model choice. Labels, class names and preprocessing (stretch vs. letterbox resizing) changed our results more than switching architectures.
How to export models to ONNX and run them in the browser with WebAssembly.
3D on the web with three.js, from a moving conveyor belt to a height map of a scan.
Good UX for high-stakes work means being calm and clear, not flashy. Showing uncertainty honestly builds more trust than pretending to be sure.

What's next for GILGAMESH

Train on larger, more varied X-ray datasets, and add more threat types (guns, batteries, liquids over the limit).
Use dual-view X-rays (top and side), the way real airport scanners do, for better 3D reconstruction.
Learn from officer feedback, so every Clear and Hold decision improves the model.
Add a real database and officer accounts.