I start with the task, source data, acceptable errors and human owner. Then I choose the model and build the assistant or agent into the tools people already use. I take responsibility for architecture, implementation, code review, tests, browser behaviour and the decision to ship.
At a glance
Put AI into a defined job, with someone accountable for the result.
- Best for
- Founders and teams with an AI use case to ship, an experiment to make reliable, an agent that needs production controls or software that needs to be shipped, tested and left maintainable.
- Commitment
- Project based, from one month, usually part time.
- First result
- In the first month: one job built, tested against its failure cases and released.
The job is narrow
It reads only approved data and does what we wrote down. It is not a general chatbot.
The error rate is set first
Before I choose a model, we agree the error rate the job can tolerate.
A person stays in charge
A named owner reviews results, handles exceptions and can stop the tool.
The team can maintain it
Tests, code review and a runbook come with the build and stay with the team when I leave.
When this fits
- People use AI ad hoc and the result varies by person.
- An experiment helps, but still relies on manual copying or prompting.
- You need an agent for a repeated task with known sources and a named owner.
- You need software shipped, tested and left in a state another engineer can maintain.
What I build
- A ranked use-case decision based on value, feasibility, source data and risk.
- Agents for sales, marketing, support and operations with limited permissions and human escalation.
- Copilots, retrieval, reporting and task automation inside existing tools.
- MVPs, internal tools, dashboards and CRM-connected workflows.
- Marketing websites that need better structure, performance and publishing workflows.
- Existing codebases that need a senior engineer to understand, refactor and extend them.
- Builds in Rails, Python and JavaScript, with API integrations.
- Monitoring for quality, cost, time saved and failure rate on everything I release.
Good agent candidates
- Research and content-brief assistants using approved sources.
- Reporting agents that prepare a first analysis from known data.
- CRM agents that summarise, route or draft sales follow-up.
- Support and operations triage with human escalation.
What agents are not for
- Legal, financial, medical or employment decisions without human review.
- Work that relies on unowned, conflicting or stale source data.
- Arithmetic, policy checks and audit trails that belong in deterministic code.
- Tasks the team itself cannot explain or assign to an owner.
What improves
- Faster scaffolding, routine code and repetitive changes.
- More time for architecture, failure cases and review.
- Repository instructions that preserve decisions, commands and conventions.
- Tests, browser checks and regression review throughout delivery.
What does not change
- I review every diff before it ships.
- Production code still needs tests, observability and maintainable patterns.
- The user problem and acceptance checks decide what gets built.
Where this connects
Related work
- Student Pay
- Soundspace Studio
- This website
- Claude Code guides
- Applied AI writing
- AI readiness writing
- Organic Growth and AI Search
Common questions
When is a team ready to move AI from an experiment into day-to-day work?
When AI already helps with a real task, the inputs and acceptable output can be written down, and the team can measure time, quality, cost or risk. It also needs a named owner who can review failures. Without those conditions, it is still an experiment.
How do you decide which AI use cases are worth building?
I rank candidates by value, feasibility, source data and risk. I start with a task people already do, where the result can be checked and the gain can be measured. I also price the maintenance and running cost before committing to a pilot.
What does an AI agent developer do?
An AI agent developer builds software that uses a language model to complete a defined task: gather inputs, use approved tools or data, apply rules and return a result for review. The engineering work includes permissions, evaluation, logging, failure handling and escalation, not just the prompt.
What makes a good candidate for an AI agent?
A repeated task with recognisable inputs and outputs, approved source material and a person who can review the result. The gain must also be measurable in time, quality, cost or risk. If the task or owner is unclear, I do not recommend an agent.
What kinds of agents do you build?
Research and content-brief assistants, first-pass reporting agents, CRM agents that summarise or route follow-up, and support or operations triage. Each agent has approved sources, limited permissions and a named owner.
What should you not use an AI agent for?
Do not give an agent unreviewed legal, financial, medical or employment decisions. Do not use one to compensate for poor source data or an undefined process. Keep arithmetic, policy checks and audit trails in deterministic code.
How do you keep a custom agent safe to run?
I restrict its sources, tools, permissions and actions; log what it does; test normal, edge and failure cases; and send low confidence or high-impact decisions to a person. Once live, I monitor quality, cost and failure rate.
What kind of software projects is this for?
Marketing websites that need better structure and performance, MVPs, internal tools, dashboards and CRM-connected workflows, and existing codebases that need a senior engineer to understand, refactor and extend them. I build in Rails, Python and JavaScript, with API integrations.
How is this different from a no-code or prototype build?
I deliver production code in a maintainable repository, with architecture, tests, observability and a clear handover. Agents help produce the code; they do not replace engineering review or production ownership.
Do you build around one AI provider or model?
No. I define the task, data, acceptable error rate, latency and cost first, then choose the model that meets those constraints. Keeping provider-specific code behind an interface makes later changes cheaper.
Do you use AI to build the software, and does that affect quality?
Yes. I give agents bounded implementation and investigation tasks, then review the diff, run the tests and exercise the rendered behaviour myself. Generated code remains a draft until it passes that review.
Who is accountable for the finished software?
The user problem decides what gets built. I remain accountable for the architecture, code review, tests, browser behaviour, release decision and maintainability of the result.
Contact
Have an AI use case that needs a technical owner?
Send the task, users, source data and consequence of a wrong answer. I will tell you whether to build it and what controls it needs.