AI Agent Developer for Hire
I build agents that take real actions in your systems, with the guardrails that decide what they may do on their own and what waits for a person.
133 completed contracts, 106 five-star reviews and a 100% Job Success Score on Upwork. 71% of that work came from clients who hired me again.
Where these projects go wrong
An agent that only talks is a chatbot. An agent that acts touches your CRM, your invoices, your customers. The moment it can act, the interesting questions are not about models: what is it allowed to do unsupervised, what does it do when a tool fails, and how do you find out what it did last Tuesday.
Most agent projects skip that layer and end up either useless, because a human re-checks everything, or unsafe, because nobody does.
What I build
Tool-using agents
Agents that read and write in your actual systems through typed tools, with retries, timeouts and clear failure behaviour.
Human approval gates
The steps that spend money, contact a customer or change a record pause for a person, with the proposed action shown in full.
MCP servers and scoped access
Model Context Protocol servers that expose exactly the operations an agent may perform, scoped and authenticated, so capability is a deliberate grant.
Run logs you can replay
Every run recorded with its inputs, tool calls and outputs, so a wrong action can be explained and reproduced rather than guessed at.
How the work runs
- Map the decision, not the chatWe write down the steps a competent person takes today, including the exceptions. That becomes the agent's scope.
- Draw the safety lineTogether we decide which actions are autonomous and which need approval. This is a business decision, and I will not make it silently.
- Build tools before promptsReliable, typed tools with sane errors do more for agent quality than prompt engineering does.
- Ship with observabilityLogging, run replay and a kill switch on day one, not after the first incident.
Proof
Multi-agent pipeline with human approval
Scout, writer, reviewer, approver and editor, each with one job. The approval step is code, so nothing is sent without a tap.
Django at scale, 21 contracts for one client
Delivery on shared standards, repeatedly, which is the same discipline agent tooling needs.
Serverless pipelines on AWS
Queues, retries and failure handling in production since 2016.
Stack
OpenAI, Claude, LangChain, LangGraph, MCP, Python, FastAPI, PostgreSQL, n8n, Make, Docker
Questions clients ask
How is this different from an automation in n8n or Zapier?
A fixed automation follows the same path every time and breaks when reality varies. An agent decides which path to take. Where a fixed workflow is enough, I will tell you, because it is cheaper to run and easier to debug.
What stops it doing something expensive or embarrassing?
Approval gates on the actions you nominate, scoped tool access so the agent physically cannot reach what it should not, and a run log so anything unexpected can be traced.
Do we need a specific model provider?
No. The tools and the workflow are the durable part. Models get swapped as they improve, and the system is built so that swap is a configuration change.
Can it run inside our infrastructure?
Yes. Self-hosted deployment with your own keys and your own database is normal, and often required for anything touching customer data.
Tell me what is broken
Send the workflow, the system or the idea. You get a straight answer on whether it is worth building, and what it would take. If I am not the right person, I will say so.