AI agent development
Agents that finish the work.
Most agent demos fall apart the moment real data hits them. I build LLM agents that use tools, pull the right context, and finish the task, with evals so you can trust what they do.
It completes the task
The agent does the whole job, not the first two steps. It calls your tools, handles the messy cases, and knows when to stop and ask a human.
You can see why it acted
Every run is traced. When something goes wrong you get the tool calls and the reasoning, not a shrug and a retry.
Tested before it ships
I write evals against your real cases so a prompt change does not quietly break the thing that was working last week.
Real tool use, not just chat.
A chatbot answers. An agent acts. I connect the model to your systems with typed tool definitions, give it the context it needs through retrieval, and put checks around the actions that matter.
- Tool calling wired to your real APIs and data
- Retrieval so the agent works from your facts
- Human approval on the steps that carry risk
The models
OpenAI and Anthropic, chosen per task on cost and quality rather than whichever one is trending.
Tool use
Typed tool definitions against your APIs, with retries and error handling for when calls fail.
Retrieval
Your documents and data indexed and pulled in at the right moment, so answers come from your context.
Evals
A test set from your real cases, run on every change, so you know quality before you ship it.
Agents I build often.
A few common shapes. We scope yours to the actual job on the call.
Support triage
Reads incoming tickets, classifies and routes them, drafts replies, and hands the hard ones to a person.
Workflow automation
Plans a multi step process, calls the right tools in order, and recovers when a step fails.
Research and reporting
Gathers from your sources, checks the facts against retrieval, and writes a report you can actually use.
Tell me the task to automate.
Fifteen minutes and I come back with a scope, a price, and a date. Worst case, you leave knowing whether an agent is even the right tool.