What I take on
AI Bots & Agent Harnesses.
Built by me, start to deploy
Support and sales bots that answer from your own documents, with the evals and the handoff that stop them inventing.
Why this gets built
A bot that invents an answer about a drawdown rule costs more than the support hour it saved. The engineering worth paying for is the refusing and the handoff.
What I build
- Answers grounded in your documents, with the source attached so staff can check what the bot leaned on.
- Handoff to a person on conditions you set, into the ticket queue the desk already works.
- An eval set that runs on every prompt change, so fixing one answer does not quietly break four others.
- MCP tools where the bot needs to read or do something real, scoped so it cannot reach past what that channel should touch.
- Transcripts with a correction path. Staff mark a bad answer and it becomes a test case rather than a note in a thread.
- Model routing, so the cheap model handles most turns and the larger one only sees the turns that need it.
How long it takes
About four weeks to a bot on one channel that answers from your docs and hands off. Each extra channel is roughly a week.
The evals are why this is four weeks and not one. Without them there is no way to tell whether a change helped or hurt.
What ships
- The bot on your first channel
- Retrieval over your docs
- Eval suite and its scorecard
- MCP tools with scoped access
- Handoff into your existing queue
- Transcript review screen
Tech stack
Next step
Let’s talk, 30 minutesNeed this built?
Tell me what you’re running and where it breaks. Half an hour is usually enough for me to say what it takes and what I’d do first.