Service
Put AI agents to work in my operation
An agent system that does real work in your business, with human approval where it matters, built by someone who runs his own company that way.
For
Founders who are spending their day on inboxes, listings, support, content or data entry that an agent with judgement could do, and who want to keep the final say.
You are here because
- You know AI could run part of your operation and every demo you have seen is a chatbot.
- You have a product and want an agent inside it that can read and change user data safely.
- Your team is trying prompts in a notebook and nothing has reached production.
- You need to know whether a cheaper model is good enough, and nobody can prove it either way.
What you get
- An agent that calls your tools, with a written brief, hard limits and an approval queue for anything outward.
- Evals: a golden set and a judge, so a model change is a measured decision.
- Cost controls: routing, caching, provider failover, a cost-per-task number you can see.
- Memory and retrieval where the job needs them, not everywhere.
- A runbook your team can operate without me.
How it runs
- One week of discovery: I shadow the work, write the brief and the limits, and we agree what the agent may and may not do.
- The first loop ships in two to three weeks, filing on an approval queue. Nothing goes out without a human click until the numbers say it can.
- We measure, widen the autonomy, and I hand it over with the evals and the runbook.
What it costs
USD 75 to 110 an hour, or a fixed price per loop
A first production loop (one job, one approval queue, evals) usually lands between USD 6k and 15k.
Proof
Kai, the AI coach inside Livio
Kai runs every coaching conversation: it streams replies, calls 30+ tools to read and change a member's data, remembers each member over time, and acts for them with one-tap confirmation. Live on both stores, kept honest by evals and cost controls.
Kata Jaam
2,451 registered users, first on Google for "events in Kathmandu", ~196 search clicks a day, the first paid placements sold in month eight. One person, with agents.
Questions I get
- Which models?
- Claude by default, through the Anthropic API or AWS Bedrock, with OpenAI where a job calls for it. I have run both in production and will show you the eval that chose.
- Will it do things without asking?
- Only what you have approved in writing, after the numbers show it is safe. Kata Jaam's agents still cannot publish a post or message an organizer without a human click, eight months in.
- Can you work inside our product?
- Yes. Kai lives inside a shipped app and changes real health data. The patterns for doing that safely are the ones I will bring.