Cobalt AI Support Desk
An AI copilot for support agents, grounded in Cobalt's own docs.
How we charted it
The challenge
Cobalt's team was drowning in repetitive tickets, but a naive chatbot risked confident, wrong answers that would erode customer trust.
Our approach
- 01
Built a retrieval pipeline over Cobalt's knowledge base with pgvector, so answers cite real sources.
- 02
Wrapped the model in an eval harness with tracing and cost controls before anything shipped to customers.
- 03
Kept a human in the loop: the copilot drafts, agents approve.
The outcome
Ticket volume dropped 41% as the copilot handled the repetitive long tail — and answer quality went up because every reply is grounded.
Why this stack
pgvector for retrieval right next to the data, a Python eval/tracing layer, and Claude for grounded, reviewable answers.
Have a build in mind? Let's chart it.
Tell us where you're headed. We'll reply within one business day with a clear, senior take — no sales theatre, no obligation.