An agentic support layer that answers before a human sees the ticket.
The client's support queue was growing faster than their team. We built a production RAG agent that now closes the majority of tickets on its own.
Support headcount couldn't keep pace with growth
The client's ticket volume had tripled in a year, and most of it was repetitive: account setup, reconciliation errors, API integration questions already answered somewhere in their docs, if the customer could find it.
A first attempt at a support chatbot had shipped and quietly been disabled after it gave confidently wrong answers on billing questions. the client needed something accurate enough to trust with real customer accounts, not a demo.
Building an eval harness before the agent
Before writing a single prompt, we built a 150-ticket held-out test set graded by the client's own support leads, so every version of the agent had a real pass rate, not a vibe check, before it touched a live customer.
The agent itself runs a hybrid retrieval pipeline over docs, past tickets, and account data via tool calls, with a confidence threshold that hands off to a human the moment it isn't sure rather than guessing.
What's in production now
The agent runs on a fused dense/BM25 retrieval pipeline with a reranking step, backed by tool calls into the client's account and billing systems for anything that needs live data rather than static docs. Every low-confidence response routes straight to a human, logged for the next eval cycle.
Four clients, on camera.
Why he brought his development work to Code Elevator.
A headless Hydrogen storefront rebuilt route by route.
A headless rebuild for a storefront carrying a heavy ad budget.
A real-time dispatch platform that replaced spreadsheets and phone calls.