Custom AI agents that finish the job
I build custom AI agents with tool calling and multi-step workflows, connect them to your systems through MCP, and debug the ones that loop, stall or cost too much.
Who this is for
- Startups that need an agent wired into their own APIs, data and tools
- Operations teams with a repeatable process worth automating
- Teams whose agent works in the demo and fails on real work
What I do
- Custom AI agents scoped to one job they can finish
- Tool calling and multi-step workflows with checks between steps
- MCP integrations with the systems you already run
- Business process automation with a human approving the risky actions
- Agent debugging and optimization: loops, wrong tool calls, latency and cost
Proof: products I shipped
- PavlySaaS · founder build
Small shops sell through reels and DMs, then track orders by hand.
- UxlyzeSaaS · AI analysis
Teams guess why landing pages do not convert.
- ZippyOpen source · dev tool
Handing tickets to coding agents means babysitting terminals and trusting unverified diffs.
How an engagement runs
- 01
Discovery. I read the code, the data and the bill, and talk to the people with the problem. You get a short plan: what is broken, what ships first, what waits.
- 02
Engineering. Working software on your real data in days. Fix what is broken, build what is missing, replace what will not hold.
- 03
Integration. Wire it into what already runs: APIs, databases, auth, payments, CRMs, MCP servers.
- 04
Deployment. Your cloud, your accounts. CI/CD, environments, monitoring and backups, so launch day is not the first deploy.
- 05
Support. Docs, handover and iteration. I own the outcome after launch, not just the merge.
Related writing
- Why AI Agents Fail in Production and How to Fix Them
Eight ways AI agents break once real users and real data arrive, how to recognize each one, and the fix that holds.
- How to Take an AI-Built App from Prototype to Production
A working checklist for apps built with AI coding tools: secrets, data access, config, payments, background work, deploys and monitoring.
- Why AI Agents Fail in Production and How to Fix Them
Eight ways AI agents break once real users and real data arrive, how to recognize each one, and the fix that holds.
- Why I built Zippy: tickets in, verified work out
Running Claude Code agents on real tickets, each in its own git worktree, and why the review step matters more than the agent.
- How Uxlyze audits a landing page with headless Chrome and Gemini
The architecture behind an AI UX audit tool: a Go worker, chromedp, a job queue and structured LLM output.
Common questions
- Which models and frameworks do you use?
- The model that fits the job and budget, usually Claude, GPT or Gemini, called through their own SDKs. I add a framework only when it removes real work.
- How do you know an agent is working?
- Saved real inputs become test cases, outputs are structured so code can check them, and every run is logged. In Zippy a second agent reviews the first and risky actions wait for a human.
- Can you fix an agent someone else built?
- Yes. I start from the logs: where it loops, which tool calls fail and what each run costs, then fix the cause instead of adding another prompt.
Other services
Have an AI product to ship?
Tell me what needs to be built, fixed, integrated or deployed. You work with me directly, the engineer who reads the code and ships it.