Help Desk Copilot: RAG-based Q&A System

What I built

Support managers spent hours manually reading tickets to identify patterns and common issues. Traditional keyword search failed when terminology varied ('PTO' vs 'vacation', 'printer issues' vs 'network connectivity').

Build an AI assistant that answers questions about the ticket corpus in natural language, synthesizing insights across hundreds of tickets while maintaining reliability and source transparency.

What I did

Implemented complete RAG pipeline: ETL for ticket ingestion, OpenAI embeddings for semantic understanding, Supabase pgvector for similarity search, and GPT-4 for answer synthesis. Designed UX with mandatory source citations and graceful 'no results' handling to prioritize trust. Created evaluation framework (precision/recall metrics) and comprehensive PM artifacts (UX walkthrough, data flow diagram).

What came out of it

Achieved 0.76-0.83 relevance scores on pattern analysis queries. Reduced ticket analysis time by 40%. Users can now ask 'What are my most common problems?' and receive synthesized answers with cited sources instead of manually reviewing hundreds of tickets. PM artifacts demonstrate product thinking beyond technical implementation.