While Large Reconstruction Models (LRMs) (Meshy or CSM) are impressive, they often produce "mushy" topology—blobs of polygons that look good but are less apt for precision manufacturing.
Mechy doesn't just "guess" the shape, but mathematically constructs it- not as a pixel prediction task, but as a logic puzzle—decomposing images into Constructive Solid Geometry (CSG) operations (Add Cube, Subtract Cylinder, etc.) to create clean, manifold, printable parts.
It weaves Gemini 3 ‘s large context window, reasoning, multimodal, coding capabilities into one agentic system with Gemini 3 Pro as the "Brain", Flash as the "Eyes" in a Self-Correcting Feedback Loop.
1. Architect (Gemini 3 Pro):
- We use thinkingConfig (Budget: 4096 tokens). It doesn't just guess but "thinks" through spatial mathematics, determining origin, scale, and CSG tree structure before producing JSON geometry.
2. Critic (Gemini 3 Flash):
- It acts as a discriminator - compares Goal (Blueprint Image) vs. Actual Result (Render Snapshot) and outputs structured feedback (e.g., "cylinder is too short").
3. Fixer (Gemini 3 Pro):
- It takes Critic's feedback and Architect's original code, uses reasoning to calculate the delta (e.g., "If it's too short, increase height property by 20%") and updates JSON.
Log in or sign up for Devpost to join the conversation.