EcoFix: The AI Repair Companion

EcoFix uses Gemini 3 to diagnose your daily broken items via live video, voice, and image upload. It provides step-by-step repair guides and calculates real-time E-waste diversion and cost savings.

Comment

Inspiration

Every year, the world generates over 50 million tons of electronic waste. Most of this comes from small devices—phones, controllers, toasters—that are discarded simply because diagnosing the repair feels impossible for the average consumer. We wanted to build a bridge between "broken" and "fixed" that feels less like a manual and more like a sci-fi companion.

What it does

EcoFix is a multimodal web application that turns any user into a repair technician. Multimodal Diagnosis: Users can upload images or use the Live Vision Camera to show the device. Voice Interaction: Users can speak natural language descriptions (e.g., "The screen flickers when I fold it") which are fused with visual data. The Impact Engine: Unlike standard chatbots, EcoFix calculates the E-Waste Diverted (kg) and Money Saved ($) for every specific repair, gamifying sustainability. Context-Aware Chat: After the initial diagnosis, users can ask follow-up questions about specific tools or steps, maintaining context of the original image.

How we built it

We built the backend using Python and Flask, acting as a lightweight wrapper around the Google GenAI SDK. The Brain: We utilized Gemini 3 Flash Preview for its speed and multimodal capabilities. The model is prompted to output structured JSON data, allowing us to parse repair steps, safety warnings, and impact metrics directly into our UI. The Frontend: The UI features a custom "Galaxy" theme built with vanilla CSS3 and HTML5 Canvas, featuring a particle system that simulates a spiral galaxy. Real-time Interaction: We used the Web Speech API for voice-to-text and HTML5 Video for the live camera feed, capturing snapshots that are sent to Gemini for instant analysis.

Challenges we ran into

Hallucinations vs. Safety: Early iterations would suggest fixes for perfectly working devices. We engineered a "Safety First" system prompt that forces the model to verify damage before suggesting repairs. JSON Consistency: Getting the LLM to consistently output valid JSON for our "Impact Cards" (Money/E-waste) required significant prompt engineering and error handling in the Python backend. State Management: implementing the "History" tab required creating a lightweight local JSON database to persist session data without a heavy SQL backend.

Accomplishments that we're proud of

The Visuals: The dynamic Galaxy theme with the rotating accretion disk animation makes the app feel futuristic and engaging. The Speed: Leveraging Gemini 3 Flash allows the "Live Capture" feature to return a diagnosis in seconds, making the experience feel truly real-time. The Logic: Successfully calculating "E-Waste Diverted" adds a tangible layer of purpose to the application.

What we learned

We learned the power of multimodal prompts. Passing an image and a user's spoken voice transcript simultaneously to Gemini 3 yielded significantly higher accuracy diagnoses than image alone. We also learned how to handle large context windows to allow for the follow-up chat feature.

What's next for EcoFix: The AI Repair Companion

AR Overlays: Using Gemini's object detection to project screwdriver turning directions directly onto the live camera feed. Marketplace Integration: Automatically linking the "Tools Needed" list to local vendors or 3D printer files for spare parts.

Share this project:

Updates