ReBrick AI — Kaggle Google DeepMind Hackathon Top 50

ReBrick AI is where it all started.

🔬 The Competition

Google DeepMind hosted "Vibe Code with Gemini 3 Pro in AI Studio" on Kaggle. It ran for exactly one week, December 5-12, 2025. The task: build an app in AI Studio using Gemini 3 Pro that solves a real problem.

4,097 teams entered in just one week. Google themselves called it "one of the largest hackathons we have ever hosted." Judging involved 30 Googlers (engineers, PMs, DevRel, marketing, design, research) over three months.

🏆 Result

Selected as Top 50. Top 1.2% out of 4,097 teams. Prize: $10,000 in Gemini API credits.

Honestly didn't expect to be picked from a pool of over four thousand. Got lucky.

🧱 ReBrick AI

The submission was ReBrick AI. This is actually where BrickQuest originated. The LEGO idea started here, and was later developed into a full-stack app for the Google Cloud hackathon as BrickQuest.

The core flow:

ReBrick AI was built with AI Studio's Vibe Coding. After confirming the potential with this prototype, I later developed it into BrickQuest with a full Next.js + Three.js + Cloud Functions stack. Same idea, completely different implementation.

💡 Two Wins

One idea, recognized twice.

- Kaggle Google DeepMind Hackathon (Dec 2025) — Top 50 out of 4,097 teams. Where it started

- Google Cloud AI Hackathon Vol.4 (Mar 2026) — Runner-up out of 260 teams. Evolved from ReBrick AI

Same problem ("what can I build with the LEGO lying around at home?"), evolved from AI Studio prototype to full-stack service, and both competitions recognized it. I think the idea itself resonated — every household with kids has the same frustration.

Judging criteria were different too. Kaggle weighted Impact (40%), Technical Depth (30%), Creativity (20%), Presentation (10%). Google Cloud focused on live pitching and Q&A. Different approaches, but both were asking the same thing: "Does this solve a real problem?"

📊 What Stood Out from the Judging

The Kaggle results announcement was interesting. Google summarized common traits among winners:

- Utility — tools that solve real friction, not just cool demos

- Agentic workflows — AI reasoning through multi-step tasks autonomously

- Native multimodality — processing images, video, audio as primary inputs

ReBrick AI didn't check all three boxes, but it touched on "real problem solving" and "multimodal" (photo → brick recognition) at least.

🎯 What's Next

$10,000 in Gemini API credits means I can use them for BrickQuest development. Improving brick recognition accuracy requires a lot of API calls, and now I can experiment without worrying about cost.

Started as a prototype on Kaggle, evolved into a full-stack app for Google Cloud, and both recognized the idea. The conviction to turn this into a real service has grown. Two competitions saw something here. Time to actually build it.