Live, 100+ users
Hinglish Mobile Repair Chatbot
Built for Mooxy
A paid freelance RAG chatbot that answers repair questions in Hinglish.
The context
Students at a mobile repair coaching institute need answers to very specific hardware faults, in the language they actually use. Generic LLM answers aren't reliable enough for that, so every reply is grounded in the institute's own repair knowledge.
What I built
- Split expert knowledge into 15+ fault domains, including charging, display, network and audio, each with its own knowledge base and prompt
- Combined FAISS vector search with re-ranking to find the right context
- Added a fault ontology and pin mapping so answers point to the right component
- Built a FastAPI and MySQL admin API so the knowledge can be updated without changing code
- Worked on it from 2025 into 2026, through local LLMs, quantised models and production hosting
The hard parts
One knowledge base wasn't enough
A charging fault and a network fault need very different context. I split the knowledge into 15+ fault domains, so each question only searches material that can actually answer it.
Repair answers have to be exact
Students work on real hardware, so vague answers are no use. A fault ontology and pin mapping tie each symptom to specific components.
Running an LLM within real limits
I worked through local models and moved to a quantised model for production hosting.
Knowledge that keeps changing
A separate FastAPI and MySQL admin API lets the content be updated and the search indexes rebuilt without touching the code.
The result
Live at Mooxy as a paid freelance project, used by 100+ people.