Bricked | Devpost

Inspiration

We love building Legos, but oftentimes our builds end up collecting dust on top of the shelf. What if we could use our existing pieces and switch up our own designs?

What it does

Bricked turns a sentence into a 3D lego model. You watch a team of AI agents plan, execute and validate the design. A planner slices it into layers, a builder fills it with different lego shapes, and an inspector physics check each step. The model stream it's building process live, with an option to jump in and steer the AI towards your vision. You can also use natural language to tweak and modify existing models.

How we built it

Multi-agent orchestrator (agents.py). Planner -> per-layer Builder -> deterministic Inspector -> Vision Critic, building bottom-up one layer at a time.
BrickGPT grammar (hxw (x,y,z)) on a 20×20×20 grid, with every placement is lintable.
Claude as the reasoning engine (planner + builders on Sonnet, vision critic on Opus), driven via the CLI.
Deterministic verification with linting (bounds/collision), connectivity, and a real force/torque physics model + a lenient "will it stand" check.
Live web app (Next.js + Three.js). SSE streams the agent tape and the geometry, an LDraw renderer assembles the model brick-by-brick, a joints overlay, and a steer bar.

Challenges we ran into

An LLM doesn't generate well in 3D. Bricks start floating, pieces don't mesh together perfectly. We solved this by backing them with BrickGPT grammar that we can lint and deterministically check.
Creating a strong physics model used to verify if our designs will sstand

Accomplishments that we're proud of

Genuine agent collaboration. The inspector can reject a layer and guide the builder back to fix.
Vision in the loop. The AI critiques its own output.
You watch it think in 3D, with the option to steer it.
Exports a working URL to buy all the components

What we learned

Some tasks are inherently not one-shot promptable. We had to decompose our task into subagents and provide a harness for interfacing with the Lego. We make harnesses and scripts to ensure that the output of an agent will be guaranteed to be correct, with no floating geometry or physics violations.

What's next for Bricked

Speed. Increasing the generation speed would help people iterate faster. Sharper shapes and being able to find better brick options that can have finer details and curves. Build with bricks you own - scan your pile and design with it.