Winning the Toyota FS + NVIDIA Tracks at HackUTD

HackUTD gives you 24 hours. We spent ours building Toyotron, an AI car-buying assistant, and it ended up taking both the Toyota FS and NVIDIA sponsor tracks. The build writeup and screenshots live on the project page, so I'm not going to repeat that here. This is about the decisions we made under the clock, because that's the part I actually reuse.

the bet: nemotron as the orchestrator

The usual hackathon setup is one big model behind a chat box and a pile of glue code deciding what happens next. We went the other way. Nemotron was the reasoning model and the orchestrator, routing work between four small sub-agents: intent, research, finance, report generation.

That one choice carried the weekend. With the orchestration inside the model, each sub-agent stayed small enough to build and test in an afternoon. Conversation state, scheduling, email generation all hung off the same spine instead of growing their own plumbing. And when something broke, we knew exactly which agent to open. In 24 hours, an architecture that tells you where the failure is beats any single feature.

The rest of the stack was boring on purpose. Next.js and TypeScript up front, a modular backend exposing tool endpoints. Boring is fast, and boring demos well.

the trade-in flow won the demo

If one feature won it, it was the trade-in flow. You upload a photo of your car. A vision model reads it, we pull VIN metadata from the NHTSA api, and a condition score drops into the same recommendation pipeline everything else feeds. Photo in, grounded estimate out, all in one conversation.

Judges don't remember agent diagrams. They remember pointing a camera at a car and watching the system reason about it. One multimodal input plus one real public data source turned "we built an agent framework" into "watch it price this actual car."

what the clock taught me

• Build the riskiest integration first. The vision + NHTSA + scoring chain was the most likely thing to die, so it went in before any polish. If it had failed, we'd have known with time left to re-scope.
• A feature you can't demo doesn't exist. Everything we kept had a visible moment in the pitch. Everything else got cut, no ceremony.
• Extensible beats exhaustive. Judges can see where an architecture goes next. They can't see the feature you half finished.
• The demo does the talking. The deck exists to get out of the way.

Two tracks with one build wasn't double the work. It was one architecture bet that happened to satisfy both briefs at once. The submission's on Devpost if you want to see the full thing.