Case study / 2025
AI Support Copilot
An agent that drafts, cites and escalates support replies.
- Role
- Fullstack and AI integration
- Client
- SaaS client via Simform
- Timeline
- 3 months
- Year
- 2025
01The problem
Support agents spent most of every ticket searching docs and old tickets. The client wanted AI help without hallucinated answers ever reaching a customer.
02What I owned
MCP server design
Agent loop and evals
Review UI and rollout
03The solution
I built an MCP server over the knowledge base and ticket history, an agent loop with Claude that drafts replies with citations, and a review UI where humans approve, edit or escalate.
Every tool call is logged. A nightly eval suite guards against regressions when prompts or models change.
04Impact
52%
Faster first response
100%
Replies with citations
3
Teams rolled out to
05Stack
- Next.js
- TypeScript
- Node.js
- PostgreSQL
- Prisma
- Claude API
- MCP
