Agentic AI engineering
I build the systems agents run on
Most of the interesting work in agentic AI is not model work. It is deciding what an agent is allowed to touch, what happens when it is wrong, and how you reconstruct afterwards what it did. I have built that layer, shipped products on top of it, and used coding agents to do the building.
MCP servers and tool design
I built the MCP server behind Morning Metrics, which lets a marketer point Claude at their own advertising data and ask a question in plain language. The design problem was not the protocol. It was deciding that the agent reads summary tables and nothing else, so there is no route from a question to a credential.
Morning Metrics, chapter 05Authorisation and guardrails for agents
That server runs on OAuth 2.1 with PKCE behind a consent screen that names what the agent will be able to read. Grants are revocable and re-checked on every request rather than once per session. Rate limits fail closed, and every tool call leaves an audit event carrying the workspace, the grant and the question.
Morning Metrics, chapter 05Prompt injection, treated as the default case
The one thing an attacker can reliably write into a marketing pipeline is ad text. Ad copy reaches the model labelled as data rather than instruction, so a headline telling the model to export the account list stays a headline. I designed for that before it happened rather than after.
Morning Metrics, chapter 05Model selection and grounding
Vegan Lens answers vegan, not vegan, or uncertain. I chose the model configuration for how it behaves when it is unsure, over a cheaper one that confidently said vegan where uncertain was correct, because a confident wrong answer is worse than no product. The facts come from open data rather than the model’s memory: the model reads the photograph and nothing else.
Vegan Lens, the model choiceShipping with coding agents
Every project on this site was built solo, with AI coding assistants doing the typing while I did the deciding. That is the part worth hiring: the architecture, the honesty calls and the decisions about what not to ship are mine, and they are visible in the write-ups rather than asserted here.
All four case studiesThe marketing side of it
I came to this from eight years of digital marketing, which is why the agent I built reads advertising data. I know what a wrong ROAS number costs, which is the reason the pipeline under it refuses to average an average or add AUD to USD.
Morning Metrics, the data layer
What I have not done
- I have not built a multi-agent orchestration system. Everything here is one agent against a tool layer I designed, which is a different and smaller problem.
- I have no evaluation harness I can show you. I tested the Vegan Lens model choice by hand, against cases I picked, and I would want to automate that before claiming it scales.
- I hold no AI certification. I am studying for one, and it goes on this site when it is real rather than while it is intended.
- None of this has a measured result attached. No user count, no hours saved, no baseline. Each case study says so in its own words rather than reaching for a number.
If you want to test any of this, test a design call rather than a claim: why the agent reads summaries instead of tables, why ad copy is data, or why the model was chosen for how it behaves when it does not know.