• 

• 

  Gemini Robotics-ER generating step-by-step assembly logic for a custom Paludarium, correctly inferring tool usage and construction order.

• 

Inspiration

We live in a golden age of software, but hardware is still gated. If a regular person has an idea—like a solar-powered cat feeder or a custom macro-pad—they hit a wall. They don't know CAD, they don't know component compatibility, and they certainly don't know how to source parts from twelve different vendors.

BuildSheet was born from a personal pain point: rebuilding a 1976 Honda CB750F. I wanted to source a very specific list—a Weisco piston kit, a Dyna S charging system, and MAC 4-2-1 headers—but I hit a gap. I needed an AI that could figure out exactly which gaskets, jets, and hardware were required to tie those high-performance parts together. I realized we didn't just need a shopping list; we needed the "GitHub of Hardware."

What it does

BuildSheet is a generative hardware platform that turns natural language into physical reality. It uses a "Consumer-to-Cloud" agentic workflow powered by a suite of Gemini models:

• The Architect Agent (Generative Design): Users describe their intent. BuildSheet uses Gemini 3 Flash to structure the project and Gemini 3 Pro to perform a real-time Technical Audit, ensuring a custom 863cc big-bore build accounts for the necessary thermal and mechanical requirements.
• The Sourcing Agent (Shopping Graph): Leveraging Google Search Grounding, the agent hunts for specific parts across the open web. It integrates with Google Maps to distinguish between "Order Online" and "Pick up at SF Moto today."
• The Kinematic Validator (Gemini Robotics-ER): Before you buy, you need to know it fits. We repurpose Gemini 3 Pro to act as a "Spatial Unit Test," generating a step-by-step Assembly Plan that calculates automation feasibility, required tools, and estimated build time.

How we built it

BuildSheet is a React-based web application powered by a sophisticated chain of specific Gemini models, each fine-tuned via system instructions for a specific engineering role:

• The Brain (Services Layer): Gemini 3 Flash handles high-speed "Drafting" loops. Gemini 3 Pro acts as the "Senior Engineer," performing feasibility audits and generating OpenSCAD enclosure specifications.
• The Logistics: We integrated Google Search Grounding to solve the "hallucination problem" by validating every suggested part actually exists and has a price.
• The Interface: The UI mimics a professional CAD tool but operates on natural language. It maintains a "Drafting Session" state that persists the BOM, generated blueprints, and assembly plans.

Challenges we ran into

The "Amazon Problem": AI is great at hallucinating parts that should exist but don't. Early versions would invent hardware that was theoretically perfect but impossible to buy. We solved this by implementing a Grounding Layer in our findPartSources service.

The "Friday Deployment Crisis" (Infrastructure Resilience): Forty-eight hours before the deadline, we hit a critical blocker when our automated AI Studio "Deploy to Cloud Run" pipeline stopped passing API key environment variables to the container. With the submission clock ticking, I had to pivot from the managed abstraction to a custom infrastructure stack. I worked with Gemini 3 Pro to audit the deployment's base image, identified a caching conflict, and architected a custom Docker solution. By manually pushing a refined base image via the Google Cloud CLI, we successfully bypassed the constraints and restored the application's agentic capabilities.

Accomplishments that we're proud of

While the commerce logic is powerful, we are most proud of the Robotic Assembly Planner. We used Gemini 3 Pro to successfully generate a valid "Temporal Execution Plan" for a complex 863cc motorcycle engine conversion. It understands the physics of construction:

• Tool Reasoning: It specifies the exact sequence for cylinder studs and identifies required tools.
• Feasibility Scoring: It calculates an automationFeasibility score (0-100), acting as a real-world sanity check for the build.
• Visual Cognition: We successfully prototyped an AR loop where Gemini 2.5 Flash analyzes a camera feed of the workspace to provide step-by-step assembly guidance.

What's next for BuildSheet

We want BuildSheet to be the "Play Store for Physical Objects."

1. Text-to-CAD: Expanding our OpenSCAD generation to render 3D models directly in the browser using a WASM-based viewer.
2. Community Marketplace: Allowing users to publish successful "Drafts" so others can order the same kit with one click.
3. Closed-Loop Robotics: Connecting our AssemblyPlan output directly to industrial robotic arms, making "Text-to-Factory" a reality.

Built With

• agentic-workflows
• ai-commerce
• aistudio
• antigravity
• developer-tools
• gemini
• gemini-2.5-flash
• gemini-3-flash
• gemini-3-pro
• generative-ai
• google-cloud
• google-maps
• google-search-grounding
• hardware-tech
• next.js
• openscad
• robotics-and-automation
• robotics-er
• typescript
• vertex-ai