RepairPal: Smart Appliance Troubleshooter

🧠 Inspiration Millions of household appliances are discarded every year due to minor repair issues, simply because users lack access to understandable troubleshooting information. Manuals are hard to navigate, repair forums are scattered, and people don’t know where to begin. We wanted to build a solution that promotes repair over replacement, reduces e-waste, and empowers users to fix common issues with AI-powered guidance.

🛠 What it does RepairPal is an AI-powered assistant that helps users troubleshoot and fix their broken appliances.

Users describe their issue in natural language (e.g., “my dryer won’t spin”).

The system retrieves relevant repair steps from:

📘 MyFixit manuals (structured JSON)

🧰 OEM repair guides (ManualsLib PDFs)

💬 Community threads (Reddit)

🔄 Fallback: live iFixit API

Results include:

DIY repair steps

Tools/parts needed

Estimated cost

Google Sheets checklist for repair tracking

Behind the scenes, it uses RAG (Retrieval-Augmented Generation), vector search, and Hugging Face models to find the most relevant fix instructions.

🧱 How we built it Data Layer:

Parsed structured repair guides from the MyFixit Dataset.

Embedded repair steps using Hugging Face’s all-MiniLM-L6-v2.

Stored documents and embeddings in MongoDB Atlas Vector Search.

RAG Engine:

Used LangChain with Hugging Face (flan-t5-base) as the LLM.

Implemented fallback fetching from the iFixit API if local data lacks context.

Frontend / UX:

Command-line interface with optional Streamlit UI.

Google Sheets integration to export DIY checklists using gspread.

🚧 Challenges we ran into 📚 Lack of structured open datasets for appliance repair — most info was either PDF manuals or forum comments. We solved this by combining structured (MyFixit) and unstructured (Reddit, PDFs) data.

🧠 Balancing the cost vs performance trade-off of LLMs — we opted for Hugging Face open models instead of OpenAI for full offline/local flexibility.

🔍 MongoDB vector search setup — creating and querying vector indexes required careful schema mapping and tuning.

🏆 Accomplishments that we're proud of Built a zero-cost, scalable AI assistant that doesn't rely on proprietary LLMs.

Created a fallback flow where the app automatically calls iFixit’s API if no matching repair info is found locally.

Contributed to the repair economy by reducing friction for everyday users trying to fix their appliances.

📚 What we learned How to implement Retrieval-Augmented Generation using LangChain and Hugging Face models.

How to use MongoDB Atlas Vector Search to store and retrieve semantic chunks of technical manuals.

The value of combining structured and unstructured data to cover knowledge gaps.

Built With

• langcahin
• mongodb
• streamlit