How to Become an AI Engineer: Skills, Projects and Jobs

Most people become AI engineers by starting as software engineers, then learning to build products on top of foundation models: calling model APIs, adding retrieval, building agents and, above all, measuring output quality with evaluations. A degree helps but isn't the deciding factor. Shipped projects with honest evals usually are.

Definition: An AI engineer builds software applications on top of existing foundation models (such as large language models), handling prompting, retrieval, tool use, evaluation, cost and reliability, rather than training models from scratch.

The title is new and messy. Two job posts with the same name can describe very different work. So before the plan, a map.

What an AI engineer does day to day

The term was popularized by a June 2023 essay from Latent Space, The Rise of the AI Engineer, which described a new kind of software engineer who builds with foundation models without needing to train them. Chip Huyen's 2024 book AI Engineering (O'Reilly) frames the field the same way: building applications with models that are already available.

In our view, a typical week mixes four kinds of work:

Building features. Wiring model calls into a product: a support assistant, a document extractor, a coding tool.
Grounding the model. Connecting it to the company's own data through retrieval, search and tools.
Measuring quality. Writing evaluation sets and checking whether yesterday's prompt change made things better or quietly worse.
Keeping it affordable and fast. Caching, choosing smaller models where they're good enough, and watching latency.

It looks a lot like software engineering. The difference is that the core component is probabilistic. The same input can give a different answer tomorrow, so testing becomes a habit rather than a phase.

AI engineer vs ML engineer vs research engineer

These titles overlap, and companies use them loosely. Here's how we'd separate them:

Role	Main question	Core work	Typical proof
AI engineer	How do we build a reliable product on this model?	Apps, retrieval, agents, evals	Shipped apps with eval results
ML engineer	How do we train and serve our own models?	Data pipelines, training, deployment	Models in production, MLOps
Research engineer	How do we make the model itself better?	Training at scale, experiments	Papers, replications, infra

ML engineers usually own the model lifecycle: features, training, serving, monitoring. Statistics and ML fundamentals matter more. AI engineers usually treat the model as a component they don't control. Product sense and software craft matter more.

The line is blurring. Many AI engineers fine-tune small models, and many ML engineers now build on top of foundation models. If research is the pull, see our guide on how to become an AI researcher. For the wider set of choices, our AI career paths guide lays out all seven.

The AI engineer skills stack, in the order we'd learn it

Skip ahead where you're already strong. The order matters more than the speed.

1. Software fundamentals. Python first. The Stanford AI Index 2026 economy chapter, using Lightcast data, lists Python as the most requested specialized skill in US AI job postings in 2025. Add APIs, Git, testing, SQL and one cloud platform. Weak fundamentals tend to show up in interviews fast.

2. Model APIs and prompting. Structured outputs, system prompts, function calling, streaming, handling rate limits and errors. Learn at least two providers so you understand trade-offs rather than one vendor's quirks.

3. Retrieval. Chunking documents, embeddings, vector and keyword search, reranking. Most real business apps need the model to answer from company data, and retrieval quality often matters more than the model choice.

4. Evaluations. The skill that separates hobby projects from production work, in our view. Build a test set of real inputs with expected outputs, score automatically where you can, and review a sample by hand. Gergely Orosz's Pragmatic Engineer reporting on AI engineering in practice (2025) points the same way: one engineering leader he interviewed named evals and non-deterministic outputs as the biggest adjustment for most developers.

5. Agents and tool use. Letting the model call tools in a loop. Anthropic's engineering guide "Building effective agents" (December 2024) found that the strongest agent builds it saw leaned on small, combinable patterns instead of heavy frameworks, and it suggests reaching for more complexity only once a simpler design has clearly fallen short. We think that's sound advice for anyone building a portfolio, too.

6. Production concerns. Cost per request, latency, caching, logging, guardrails, and handling sensitive data.

7. Enough ML to reason about models. How training and fine-tuning work, what embeddings are, why models hallucinate. You don't need to derive backpropagation. You do need to know when fine-tuning is the right tool.

What the AI engineer job market data shows

No federal category exists for "AI engineer," so the data comes from neighboring sources.

The government baseline. The closest BLS occupation is software developers, who earned a median of $135,980 in May 2025. It projects the wider software developers, QA analysts and testers group to grow 10 percent from 2025 to 2035, with about 106,100 openings a year on average. For context, computer and information research scientists, a much smaller group whose work the BLS describes as including machine learning, had a median of $140,300 in May 2025. Our guide to the highest-paying tech jobs shows how those figures compare across roles.

Demand for AI skills. The Stanford AI Index 2026 found that AI-related postings made up 2.6 percent of all US job postings in 2025. Small as a share, large in absolute terms.

Startup pay. Carta's State of Startup Compensation H2 2025 found that median salary for AI and ML engineers on its platform rose 9.1 percent between January 2024 and February 2026, against 6.4 percent for individual contributors overall. Median initial equity grants for AI and ML engineers grew 31 percent over the same span. Our salary guide is another reference point.

Our read: demand is real, but it concentrates on people who can show shipped work. Generic "AI enthusiast" profiles face a crowded field.

Projects that prove AI engineering skill

A portfolio does more work than a certificate here. Three or four finished projects beat ten tutorials, in our view. Ideas with a clear hiring signal:

A retrieval app on a real corpus, such as a city's permit rules or an open-source project's docs, with an eval set and published accuracy.
An extraction pipeline that turns messy PDFs or emails into structured data, with error rates by field.
A small agent that completes a multi-step task using two or three tools, with logs showing where it fails.
A contribution to an open-source AI library, even a fix or a docs improvement with a test.

For each, write a short README: the problem, the design, the eval method, the results, the cost per request, and what you'd do next.

Worked example: an eval-driven portfolio project

All numbers here are illustrative. Say you build a question-answering bot over a software project's documentation.

Eval set: 150 real questions pulled from the project's forum, each with a correct answer you wrote down.
Version 1, prompt only: 87 of 150 correct, or 58 percent.
Version 2, with retrieval: 117 of 150, or 78 percent.
Version 3, retrieval plus reranking and a citation check: 129 of 150, or 86 percent.

Now cost. At illustrative prices of $3 per million input tokens and $15 per million output tokens, a request with 3,000 input tokens and 300 output tokens costs about $0.0135. At 10,000 questions a month, that's about $135. Trim the retrieved context to 1,200 tokens and the cost drops to about $0.0081 per request, or $81 a month, if accuracy holds on your eval set.

That README tells a hiring manager more than any course certificate: you measured, improved, and thought about money. Prices vary by provider and change often, so treat these as placeholders.

How to become an AI engineer: a step-by-step plan

One reasonable path, assuming you already code:

Weeks 1 to 2: foundations. Two model APIs, structured outputs, function calling. Build one tiny app end to end.

Weeks 3 to 6: retrieval and evals. Build project 1 above. Write the eval set before you tune anything.

Weeks 7 to 10: agents and production. Build project 3. Add logging, cost tracking and a fallback when the model fails.

Weeks 11 to 12: publish and apply. Clean READMEs, a short write-up of what you learned, then targeted applications. On the 1752vc careers board's AI track, filtering by Entry or Mid level and Past 7 days surfaces fresh AI engineering roles at AI companies and startups.

If you don't code yet, plan for longer, often a year or more, and learn software engineering first. Our pillar guide on how to get a job in AI covers the non-technical doors if you'd rather not take this route.

If you're a software engineer now, the fastest route is often internal: volunteer for the AI feature on your current team.

AI engineer interview prep

Formats vary, so check each company's own guide. OpenAI's public interview guide describes a skills assessment that can include pair coding, take-home projects or technical tests, followed by four to six hours of final interviews with four to six people over one or two days. Anthropic's careers page notes it uses live coding tools for technical roles and that you can look things up during them.

What we'd prepare for:

Coding. Standard data structures, clean Python, tests. Many loops still include it.
LLM system design. For example, designing a support assistant for 1 million customers. Cover retrieval, evals, latency, cost, failure handling and privacy.
Evals. Expect to explain how you would know a feature got worse, with a concrete method.
Debugging. A prompt or pipeline that fails on edge cases. Talk through how you'd isolate it.
Your projects. Expect deep follow-ups on trade-offs you made.

Our guides on preparing for an AI job interview and writing a resume for an AI job go deeper.

"Real AI engineers train their own models"

Some engineers argue that building on someone else's model is just API plumbing, and that lasting careers belong to people who understand training deeply. There's something to it. Foundation model APIs change fast, and skills tied to one vendor's quirks can age badly.

But most companies don't train frontier models. They need people who can make an existing model reliable, safe and affordable in a product. Evals, retrieval and systems design carry across vendors. Our take: learn enough ML to reason about models, and put most of your time into shipping.

Common mistakes

Collecting courses instead of shipping. Finished projects are easier to judge.
No evals. A demo that worked once tells a reviewer little.
Framework-first learning. Understanding raw API calls makes frameworks easier to debug later.
Ignoring cost and latency. Production teams care about both.
Applying by title only. Read the responsibilities: some "AI engineer" posts are really ML or data roles.

Where we land

We think AI engineering is one of the most accessible technical doors into AI right now, especially for working software engineers. The bar isn't a PhD. It's evidence you can turn a model into a dependable feature and prove it with numbers.

That's our view, and the field moves quickly. Check what the companies you want are actually hiring for. The AI jobs track shows each employer's own posting date, which helps you see what's current.

The bottom line

Anyone can make a model say something impressive once.

AI engineers make it say the right thing on the ten-thousandth try.

Key takeaways

An AI engineer builds products on top of existing foundation models, while ML engineers usually train and serve their own.
A practical skills order: software fundamentals, model APIs, retrieval, evals, agents, production concerns, then enough ML to reason about models.
There's no BLS category for AI engineers; software developers earned a median $135,980 in May 2025, and Carta found AI and ML engineer pay rising faster than other roles.
Three or four projects with eval sets, results and cost figures tend to prove more than certificates.
Interviews often mix coding, LLM system design, eval questions and deep dives on your own projects.

Frequently asked questions

What is the difference between an AI engineer and an ML engineer?

An AI engineer usually builds applications on top of existing foundation models, focusing on prompting, retrieval, agents, evals and cost. An ML engineer usually works on the model itself: data pipelines, training, serving and monitoring. The titles overlap, and many companies use them interchangeably, so the job description is a better guide than the title.

How much do AI engineers make?

There's no official pay series for AI engineers. The closest BLS category, software developers, had a median wage of $135,980 in May 2025. Carta's startup data shows median salary for AI and ML engineers rose 9.1 percent from January 2024 to February 2026, faster than the 6.4 percent for individual contributors overall. Pay varies widely by company, level and location.

Do AI engineers need a computer science degree?

Not always. The BLS lists a bachelor's degree as the typical entry education for software developers, and many postings ask for one. But some AI employers say they care more about demonstrated ability than credentials; Anthropic, for one, notes that about half its technical staff had no prior ML experience. Strong projects can offset a missing degree, though it may narrow your options.

How long does it take a software engineer to become an AI engineer?

In our view, a working software engineer can often build a credible AI engineering portfolio in about three months of focused part-time work: model APIs, a retrieval project with evals, and a small agent. Someone starting without coding experience should usually plan for a year or more, since software fundamentals come first. Moving into AI work on your current team can be faster still.

What projects should be in an AI engineer portfolio?

Strong options include a retrieval app over a real document set with published accuracy, a data extraction pipeline with error rates, a small multi-step agent with logs, and an open-source contribution. Each project helps most when its README explains the problem, the design, the evaluation method, the results and the cost per request.

Sources

Disclaimer: This guide is for general education only and is not legal, tax or investment advice. Laws, market data and program terms change, so it may not reflect the latest developments or fit your situation. Treat it as a starting point, not a source of truth, and talk to a qualified lawyer, accountant or financial adviser before you make decisions.