Sourced | Devpost

What it does

Sourced is an AI-powered supply chain transparency tool that allows users to upload a photo of any electronic device to instantly explore its internal components and trace their global origins. It generates an interactive 3D exploded view of the device's internals and visualizes the manufacturing journey on an interactive global supply chain map. This helps consumers understand the complexity and environmental impact of their devices, directly supporting UN SDG 12 (Responsible Consumption and Production).

How I built it

I built a modern full-stack application using React 18, Vite, and Tailwind CSS for the frontend, with React Three Fiber for the high-fidelity 3D rendering and react-globe.gl for the supply chain visualization.

I powered the backend with Python Flask to orchestrate a sophisticated hybrid AI pipeline:

Google Gemini 2.5 Pro Vision analyzes the uploaded photo to identify the specific device model. Meta's SAM 3D (Segment Anything in 3D) is utilized for initial 3D reconstruction. Gemini 2.5 Pro acts as a "procedural engine," using its knowledge of real-world teardowns to generate detailed JSON specifications for internal components (batteries, logic boards, sensors) when 3D scanning is insufficient. Gemini with Search Grounding conducts real-time research to map these components to their actual manufacturing locations globally.

Challenges I ran into

3D Generation from 2D Images: Extracting detailed 3D internal structures from a single 2D photo was my biggest hurdle. SAM 3D often only captured the outer shell. I overcame this by building a fallback system where Gemini procedurally generates the internal components based on teardown knowledge, requiring extensive prompt engineering to enforce physical constraints (no overlapping parts) and realistic geometry. Component Positioning: Teaching an LLM to "think in 3D coordinates" was difficult. I had to define strict coordinate systems and "industrial design rules" in my prompts to ensure components like batteries and CPUs were placed logically within the device chassis. Visual Fidelity vs. Performance: I wanted the app to feel "premium" with realistic materials (glass, metal, silicon). Balancing these high-quality 3D renders with the performance constraints of a web browser required careful optimization of my Three.js scene and material presets.

Accomplishments that I'm proud of

The "Magic" Factor: I am incredibly proud of the seamless "photo to exploded view" experience. Taking a picture of a phone and immediately being able to interactively "explode" it to see the chips and sensors inside feels genuinely magical. Hybrid AI Implementation: Successfully combining the spatial capabilities of SAM 3D with the semantic reasoning of Gemini 2.5 Pro. I didn't just use one model; I built a pipeline where they complement each other to create a result neither could achieve alone. Real-World Impact: I turned the abstract concept of "supply chain transparency" into a tangible, interactive experience that anyone can understand, making a complex global issue accessible and engaging.