Challenge: A SaaS company was drowning in repetitive support tickets and wanted accurate self-service without hallucinated answers.

What we built: A retrieval-augmented assistant grounded in the company's docs and past tickets, with citations, confidence gating and a clean hand-off to human agents when unsure. Every answer links its sources.

Impact: Around 60% of incoming tickets were deflected with grounded answers, and first-response time on the rest dropped sharply.

Stack: Claude API, vector search, Laravel, queue-based ingestion, Blade admin console.