Keel Agents
Send a brief
Menu

Two projects I built to show how I work

Both are portfolio projects. I designed and built them myself, to the standard I would hold client work to, and no client commissioned either. I have no client case studies to show yet, so the evidence here is the engineering: tests, measured results and the decisions behind them, stated as exactly as I can.

agent system, approval gates

AI service delivery coordinator

Turns a client request into a plan that a person edits and approves, creates the project in Asana and HubSpot, then watches delivery and proposes fixes for blockers, missing input and slipping dates. Nothing is written to either tool until someone approves.

Status
v1.0 released. The repository is private, and there is a walkthrough on YouTube.
Service
The kind of build behind my AI agents and workflow automation work.

Read the coordinator case study

project page / draft plan
Screenshot of the coordinator’s project page showing a draft plan for a bakery’s grand-opening campaign. A banner reads 9 warnings to check before approval, and buttons offer Edit plan, Approve plan and Decline project. Below, tasks are listed with their roles, durations and dates.
A draft plan a person has edited. Approving it is what releases the project to Asana and HubSpot.

~6,800tests

Unit and integration tests.

source: project README

85browser tests

CI also runs strict mypy and both Docker Compose stacks.

source: project README

RAG, citations checked in code

EU AI Act compliance assistant

Answers questions about the EU AI Act in a few sentences, each citing the provision it comes from, or refuses when the Act does not support an answer. Code checks the citations, not the model.

In my test, a correct provision was in the top five results for 0.86 of the questions with vector search alone (recall@5). The hybrid search I shipped scores 0.58. The case study says so, and shows why.

Status
Public repository on GitHub. It runs locally with your own OpenAI key.
Service
The kind of build behind my RAG development work.

Read the assistant case study

route of one question
The route of a question through the EU AI Act assistantA person asks a question. The assistant searches the Act, then code checks the answer. A passing answer is shown with its citations. A failing one becomes a refusal.QuestionSearchCode checksCited answerRefusal
A question is searched, then checked in code. The result is a cited answer or a refusal.

245offline tests

Run with no API key.

source: eu-ai-act-assistant repo

36hand-labelled questions

The retrieval eval set.

source: DECISIONS.md

0.86recall@5, vector only

The hybrid search I shipped scores 0.58.

source: DECISIONS.md

Tell me what is eating your team’s week.

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 brief

Or write to keel.agents@gmail.com.