about
Keel Agents is the name I work under: one independent AI engineer building agents, document Q&A and LLM features for small businesses, agencies and startups.
I am self-taught and came to AI engineering from backend and web apps. I would rather show than claim: each project below comes with its tests, its measured results and its limits written down. My public code is on GitHub.
projects
Two projects, built on my own to the standard I would deliver to a client. They are portfolio projects, not client engagements. Read the evidence rather than take my word for it.
You ask a question about the EU AI Act. Every sentence of the answer cites the provision it came from, and the code checks that the provision was really retrieved. When the Act does not support an answer, the assistant refuses.
It turns a client request into a project plan. A person edits and approves the plan; only then does anything reach Asana and HubSpot. After that it watches delivery and proposes what to do about blockers and slipping dates, for a person to approve.
~6,800tests
Unit and integration tests.
source: project README
85browser tests
CI also runs strict mypy and both Docker Compose stacks.
source: project README
habits
These five habits show in both projects. They are also what I would bring to yours.
Anything with consequences waits for a person. The AI can propose a plan or draft a message. It cannot send it, or write it into your systems, on its own.
evidence
In the coordinator nothing is created in Asana or HubSpot until a person approves. The approval records who approved which version of the plan.
If a claim about quality matters, a test or an evaluation sits behind it. I measure against questions I labelled by hand, and I publish the result even when it is unflattering.
evidence
On the assistant, vector search alone scored 0.86 on recall@5; the hybrid search I shipped scored 0.58. Both reached 0.92 at recall@20. The table and my reasoning are in the repository.
Anything that calls a paid model has a hard limit written into the program, not a polite request in a prompt.
evidence
The assistant stops at 10 questions per session and 30 per address. In the coordinator, every model call reserves tokens against a per-run and a daily cap before it is made.
What a customer types, a comment on a task or a note in a CRM is never an instruction. It reaches the model quoted, never as part of the instructions.
evidence
In the coordinator the agents have no tools, so a hostile comment has nothing to act with. In the assistant, every line of visitor text is quoted, so a question cannot pose as a provision of the Act.
Every project I hand over comes with a list of what it does not do. You should hear that from me before you find it out.
evidence
The assistant’s README says its citation check confirms a cited provision was retrieved, not that the sentence says what the provision says. The coordinator’s lists what is not built yet, including that it has no hosted demo.
engagements
Send a short brief: what the work is, who does it now, and what it costs you. I reply within two working days with a rough scope, a price range and whether AI is the right tool for it at all.
Send a briefOr write to [email protected].