BrickThis

#devchallenge #mlhreflections #webdev #geminireflections

Built with Google Gemini: Writing Challenge

This is a submission for the Built with Google Gemini: Writing Challenge

What I Built with Google Gemini

Problem: As a kid, I wished I could build anything with LEGO. But I wasn’t a master builder; I couldn’t figure out how to get the shape I wanted to just from imagination. Stuff fell apart, pieces got stuck, it was a mess. Instruction manuals were so clear and easy to read.

Solution: Turn any photo into a LEGO build! Snap or upload an image and get a parts list plus step-by-step instructions.

Google Gemini 3 Pro analyzes the image, outputs a JSON file containing all necessary parts (including color, quantity, and ID), instructions, and prompts for image generation. A dataset of the top 50 LEGO pieces is what Gemini uses for accuracy, along with data from the Rebrickable API. Flux then creates a cover photo and diagrams from the Gemini output.

I also used Google AI Studio to create a quick prototype of the UI as a starting point, tweaking it to my liking afterwards.

Demo

https://brick-this.vercel.app

What I Learned

This was my first Hackathon, and I learned so much from it. Thinking about how it works under the hood was the hardest part. The AI produces something that resembles a LEGO set, but it wasn’t entirely accurate in terms of parts matching one-to-one with the image generation. I could have fixed this if I had more time.

Before even touching prototyping or UI, it’s best to get the backend fully working the way you want. 3/4 into the project, I realized there were a few problems. Flux has no idea what the last step looks like, so image sequences can feel incohesive. Even though I planned it out, I wish I tested more.

Also, for some reason, I decided it was a great idea to deploy the site at the last minute, and of course: build errors. Next time, I’ll deploy the site first thing.

Lastly, I was sad I didn’t find a team I clicked with at first, but I found out I really enjoy being a solo dev. I get to be as creative as I’d like!

Google Gemini Feedback

I think the model had trouble picturing the build as a whole physically. There were a few times when the text descriptions weren't clear or detailed enough, which caused the image model to improvise.