An AI support agent that survives production.
A consumer software company needed their legacy chat bots replaced with something that could actually resolve support conversations, on the phone and in chat, without inventing answers or leaking account data.
The build
A production LangGraph orchestrator in TypeScript replaced the legacy agents: a retrieval augmented support agent with hybrid vector search over the company's docs, a compliance gated sales flow, and dedicated state machines for retention and spam routing. The same platform serves voice through a websocket adapter and live chat in the browser.
Account operations are the part most teams get wrong. Lookups, verification codes, orders and cancellations run through an MCP server over JSON-RPC, so the model never touches credentials and every action is auditable.
The discipline
Langfuse observability on every conversation, eval harnesses that catch quality slips before customers do, and a CI ingestion pipeline that keeps the knowledge base current as docs change. AI features are software. They deserve the same engineering as the rest of the system.
“Wes is a highly engaged and sharp developer who proactively analyzes systems and naturally looks for optimizations.”
Want AI features that hold up after the demo? That is the exact thing we build.
Book a call