What is an AI engineer?

The rolesCopy Anchor Link

An AI researcher invents new methods and trains frontier models. A machine learning engineer trains and deploys custom models on an organization's data. An AI engineer uses pretrained models through APIs or open weights, and builds the system around them: prompts, retrieval, tools, evals and deployment. Software engineers already have most of the skills: APIs, databases, testing and operations. What is new is working with a non-deterministic component.

What the work looks likeCopy Anchor Link

Choose a model. Design prompts and context. Add retrieval and tools. Build evals to know whether it works. Trace and monitor it in production. Control cost and latency. Keep it safe. The roadmap in this site follows that order.

AI is narrow, AGI is a goalCopy Anchor Link

All current systems are narrow in the sense that they have uneven abilities: superhuman on some tasks and brittle on others. AGI, artificial general intelligence, has no agreed definition. Do not design around it. Design around what the model measurably does on your task today.

How to growCopy Anchor Link

Build small projects that break in interesting ways. Read traces of your own system. Learn evals early. Keep a short list of sources and ignore the noise.

A worked example: a day in the life of an AI engineerCopy Anchor Link

Morning: read yesterday's failed conversations from the assistant and group them by cause. Midday: change a prompt, add a missing tool, rerun the test set to see if the pass rate moved. Afternoon: check cost and speed dashboards, review a pull request, talk to the people who use the product. Notice how little of it is "training a model" and how much is testing, reading and fixing.

What you need to startCopy Anchor Link

Basic programming, comfort reading error messages, curiosity about why something failed, and the habit of measuring. Math helps for understanding but is not required to build. This path covers the math in a gentle way so you can follow the research when you want to.

A worked example: finding which stage failedCopy Anchor Link

A support assistant tells a customer the refund window is 60 days. The policy says 30. The work is locating the stage that failed.

1. Open the trace, the logged record of the conversation and every step inside it.
2. Check whether the retrieved documents included the 30 day policy.
3. If not, the retrieval step is the problem. Fix how documents are split or searched.
4. If yes, the prompt or the model ignored it. Reword the instructions or try another model.
5. Add this case to the test set and rerun all cases, so the fix cannot quietly break something else.

Trade-offs: call an API, run open weights, or trainCopy Anchor Link

Calling a hosted model through an API is the fastest start. You pay per use, your data leaves your systems, and the provider can change or retire the model. Running open weights gives you control and privacy, but you now operate the servers. Training your own model makes sense only when you have enough data and nothing existing meets the need. Most products should start with the first option and move only when a measured problem pushes them.

What goes wrong in the roleCopy Anchor Link

Three symptoms come up again and again.

1. The demo works but production fails. The cause is usually testing on five easy examples. Build a test set of real cases.
2. A change feels better but you cannot tell. Without a score, you are guessing. Run the same set before and after.
3. Results differ between runs. The model is non-deterministic, so run each case several times.

Try it yourselfCopy Anchor Link

Take a document you own. Write 10 questions whose answers you know. Ask any chat model, mark each answer right or wrong, and note the score. Change the instructions and rescore. Expect the score to move by a case or two, sometimes up and sometimes down on identical input. That noise is why real projects need larger test sets than 10.

Common mistakeCopy Anchor Link

Waiting to feel "ready". The field changes every few months, so nobody is. Build something small this week.