AI engineering

Hire an AI developer who is a senior engineer first

Hire an AI developer through me when the feature has to work in a real product: an assistant that answers from your own data, an agent that takes actions in your system, document processing that used to be manual, or an MCP server that opens your product to AI agents. I build it in your codebase, with the evals, permissions and cost controls that production needs.

I’m an application engineer with twenty years in production. I use models from providers such as Anthropic and OpenAI; I don’t train them. Most of the work in an AI feature is ordinary engineering around a model, and that is the part that usually goes wrong.

AI work I take on

  • LLM features in existing products. Summaries, drafting, classification, extraction and search, wired into the product you already have. More in AI integration.
  • Retrieval (RAG). Answers grounded in your documents and data, with access rules respected. See RAG development.
  • Agents. Software where the model plans and calls your APIs, with scoped permissions, audit trails and a human approval step. See AI agent development.
  • MCP servers. A typed, permissioned interface so AI agents can use your product. See MCP server development.
  • Evals and cost. Test sets that tell you whether a prompt or model change made things better, and spend that is measured per feature. See LLM evaluation.

AI I’ve shipped

Craaft is a kanban board where people and AI agents work the same cards, through an MCP server or the REST API, under the same permissions. I designed and built both. The MCP server is public.

Ecko is a programming language I wrote in which ai is a keyword: model calls return typed output, tools are declared in the language, and a deterministic mock mode lets programs be tested offline.

For client teams I build LLM features into existing products and roll out AI coding tools across engineering teams.

What goes wrong without a senior engineer

  • No evals. Nobody can say whether last week’s prompt change helped, so every change is a guess.
  • Permissions bolted on. The model can read or do things the user asking couldn’t. This is the most common serious bug in AI features. See AI agent security.
  • Unbounded cost. One customer’s usage or one retry loop produces the bill. Budgets and caching belong in the first version.
  • No fallback. The provider has an outage or a rate limit, and the feature takes the page down with it.
  • A demo, shipped. It worked on ten examples. Production sends ten thousand, including the ones nobody thought of.

How a contract works

It starts with a two-week sprint: one feature, a small eval set and a working version in your staging environment. After that I work a fixed number of days a week, agreed up front. I invoice monthly from my own company, in your currency, on 30-day terms.

I work remotely from Cape Town (GMT+2), with a full working day of overlap with Europe and the Gulf and mornings with the US East Coast.

Questions

What does an AI developer do?

An AI developer builds software features on top of language models: choosing the model, designing prompts and retrieval, connecting tools and data, and adding the testing, permissions and monitoring that make the feature dependable. It is application engineering, distinct from the research work of training models.

Do you train or fine-tune models?

I don’t train models from scratch. Most products don’t need fine-tuning either; retrieval and good evals get further for less. I cover when it is worth it in RAG vs fine-tuning.

Which models and providers do you work with?

The major hosted models, including Claude and GPT, and open models where data has to stay in your own infrastructure. I build so the model can be swapped, because the best choice changes every few months.

Can I hire you as a freelance AI engineer for a single feature?

Yes. A single well-scoped feature is a good first contract, and the two-week sprint is designed for it.

Need an AI developer who ships to production?

Tell me the feature and the product it goes into. In a two-week sprint I’ll build a working version with evals, then you decide.

GMT+2 · Full EU and Gulf overlap · US East mornings · Singapore afternoons

Related