Winning the Toyota FS + NVIDIA Tracks at HackUTD

Twenty-four hours. Two sponsor tracks. One working prototype. That was the math at HackUTD, and the thing we shipped — Toyotron, an AI-powered car-buying assistant — took both the Toyota FS and NVIDIA tracks. This isn't the build deep-dive; it's about the decisions that had to be made under a countdown, and which ones I'd make again.

The bet: make Nemotron the orchestrator, not just the chat model

The default hackathon architecture is one big model behind a chat box, with a pile of glue code deciding what happens next. We bet against that: NVIDIA Nemotron ran as both the core reasoning model and the multi-agent orchestrator, coordinating four specialized sub-agents — Intent, Research, Finance, and Report Generation.

That scoping choice is the whole story. With the orchestration inside the model — dynamic task routing through prompt-layered governance, with persistent session context — each sub-agent stayed small enough to build and test in an afternoon. Conversation state, scheduling workflows, and email generation all hung off the same spine instead of each growing its own plumbing. When something misbehaved, we knew exactly which agent to open up. In a 24-hour sprint, an architecture that localizes failure is worth more than any single feature.

The rest of the stack was deliberately boring: a Next.js + TypeScript frontend and a modular backend exposing API-based tool endpoints. Boring is fast. Boring demos well.

The feature that made the demo land

If I had to point at one thing, it's the trade-in flow. Upload a photo of your car, and an OpenAI vision model runs image recognition on it, pulls VIN metadata from the NHTSA API, and fuses a condition score into the same recommendation pipeline everything else feeds. Photo in, grounded estimate out — in one conversational flow.

It worked because it's visceral. Judges don't remember your agent diagram; they remember pointing a camera at a car and watching the system reason about it. Multimodal input plus a real public data source turned "we built an agent framework" into "watch it price this actual car."

What the clock teaches you about scoping

The lessons I keep reusing from that weekend:

• Build the riskiest integration first. The vision + NHTSA + scoring fusion was the most likely thing to fail, so it went in before any polish. If it had died, 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 that didn't got cut without ceremony.
• Extensible beats exhaustive. We shipped a polished, extensible prototype — not a complete product. Judges can see where an architecture goes next; they can't see the feature you half-finished.
• Demo > slides. The prototype did the talking. The deck existed to get out of its way.

Winning two sponsor tracks with one build wasn't about doing twice the work — it was about one architecture bet that satisfied both briefs at once. The full write-up and screenshots are on the Toyotron project page, and the submission is on Devpost.