Thought Partner and Builder for Agentic Systems

TERMS Part-time and hourly. Realistically under 5 hours a week to start, growing if it works out. I run a loyalty analytics consultancy. Small team, senior people, and most of what we collectively know is sitting in people’s heads, old documents, and meeting recordings. I have been building our AI systems myself. I have got further than I expected to, and I have hit the ceiling of what one person thinking on their own does well. I want someone to think with. To be clear about the hours: this starts small. Under five a week, most likely. I would rather have a few hours from the right person than a full week from the wrong one, and if it goes well there is a lot more to do. WHAT I'M ASKING FOR Not someone I hand a spec to. Most of the work is deciding what is worth building in the first place, then getting to something usable before the idea goes stale. In practice: I bring a messy problem out of the business, and we work out together whether an agent is even the right answer and what it should look like. You tell me when I am wrong and why. We get to a working prototype fast, because arguing about something we can actually use beats arguing about something we are imagining. And you tell me when a prototype should be thrown away. If you read that and think you would rather have defined scope and a clean handoff, you would hate this. Better we find that out now. WHAT'S ON THE LIST A company brain. Somewhere our team can ask what we know and get an answer that cites where it came from, without leaving the tools they already work in. Agentic workflows for the recurring work of the business, with approval steps on anything that matters and clear limits on what runs unattended. AI-native software tools. Not a chat box bolted onto a screen that already exists, but tools designed from the start around what an agent can do. That is where CopilotKit comes in. And the one I care most about: a learning loop that makes all of the above better the more we use them, instead of starting from zero every session. If you have real opinions about how to tell whether an agent actually got better, that is close to the top of what I want to talk about. TECH STACK Python, LangChain, LangGraph, and LangSmith. That is the core and we are not looking to swap it for a low-code platform. CopilotKit is where I expect to land for agent-facing UI. If you have shipped with it, say so. It is uncommon enough that I am treating it as a signal rather than a requirement. All of this has to live inside what we already use, which means Claude and Claude Code, Codex, and the Microsoft suite, particularly Teams, Word, and Excel. Genuine Microsoft Graph or Office integration experience is rare and I would weigh it heavily. ENGINEERING DISCIPLINE Whatever we keep has to be real software: -- Unit and integration tests, with CI that runs them -- Evaluation sets and regression tests on agent and retrieval output, so a quality drop shows up before a person notices it -- Tracing and observability from the start, token cost and latency included -- Structured logging, alerting, and a way to roll back -- Enough architectural documentation that the next engineer can pick it up WHAT THIS IS NOT Not an n8n, Make, or Zapier build. Not a chatbot. Not a seat where you wait for tickets. If most of your work is wiring SaaS tools together, we are not a fit. IF YOU WANT TO APPLY Skip the general pitch and answer these: 1. Tell me about a time you talked a client out of what they asked for. What they wanted, what you pushed for instead, how it turned out. This is the answer I care most about. 2. Two systems you built and ran. What broke once it was live, and how did you find out? 3. How would you measure whether an agentic system is getting better over time? Be specific. 4. What do you hand over besides code? Tests, evals, diagrams, dashboards. Show me what working with you leaves behind. I read all of them and I will reply either way.