AI Engineering: How to Land Your Dream Job (Using the Skills Map)

Most AI engineering applications fail for the same reason: the candidate can demonstrate that they used a model, but not that they made an unreliable component behave reliably in production. Hiring teams stopped being impressed by the first thing in 2023. The second is still scarce.

This is the candidate-side version of Andrew Ng's AI Engineering Skills Map. The map tells you what to learn. This post tells you how to turn each of the four skills into evidence a hiring panel can score.

The four skills, and what each one has to prove

Work through them in that order. The first two are the entry ticket; the last two are what decides your level and your offer.

Skill 1: Build something that survives contact with reality

The differentiator Ng names explicitly is evaluation-driven development — the ability to look at traces, run error analysis, and decide what to try next. Almost nobody's portfolio shows it.

What to put in a project so it does:

• An eval set you built yourself. Fifty to two hundred real examples with expected behaviour, plus the reason each one is in the set. Say which failures it caught.
• A before-and-after number. "Grounding accuracy went from 71% to 92% after I changed chunking and added a reranker" beats any architecture diagram.
• A deliberate grounding decision. Vector index, knowledge graph, or a semantic layer over structured data — and why the others were wrong for your data.
• One deterministic guard. The place you decided not to use an LLM is a stronger signal than the places you did.
• Production traces. Even a small deployment with logging, drift checks and a cost-per-request figure.

The detail behind each of these sub-skills is in Building and deploying AI applications.

What to delete from your portfolio: tutorial-shaped RAG chatbots, wrapper apps with no evaluation, and anything whose README describes the stack rather than the problem.

Skill 2: Fundamentals are the filter, not the formality

Agents produce more code than any team can carefully read, which raises the value of the person who can tell whether it should be merged. Interview loops have adjusted: the code-review round is now more common than the algorithm round.

Be ready to:

• Take agent-generated code and name what breaks first under load — an N+1 query, an unbounded retry, a payment write with no idempotency key.
• Walk through one trade-off end to end: the options, the constraint that decided it, and what you would revisit at 10x scale.
• Describe a system you debugged that you did not write, hypothesis first, not trial and error.
• Say what you would delete from something you shipped. Strong engineers have a list.

The full rubric interviewers score you against is in Software engineering fundamentals in the age of coding agents.

Skill 3: Show how you supervise agents, not that you use them

Everyone uses coding agents now, so "I use Claude Code daily" carries no information. What differentiates is the supervision loop. Prepare a two-minute narration of a real session:

1. What context you gave the agent, and what you deliberately withheld.
2. Where you let it run, and the point at which you intervened.
3. How you verified — tests written before the agent ran, a diff you read line by line, a staging check.
4. What you refused to delegate, and why.

If a company allows AI in the interview, use it exactly the way you would on the job. Candidates who quietly paste output lose; candidates who narrate what they are checking win.

Skill 4: Shaping the build is where the offer level is decided

This is the skill hiring processes assess worst and pay most for. As agents get better at executing a spec, the value moves to deciding what belongs in it. Two answers you should have ready before any onsite:

• What you argued against building last year, and whether you were right. This demonstrates product judgment and intellectual honesty at once.
• The business metric your last AI feature moved. Not tokens saved, not latency — revenue, retention, handling time, error rate.

Candidates who can connect a technical decision to a business consequence get hired above the band they applied for. It is the same pairing the AI-First Talent framework screens for from the employer side: domain expertise and business context first, AI fluency layered on top.

A 90-day plan if you are starting from software engineering

• Weeks 1–4 — LLM foundations: tokenization, context windows, tool calling, sampling, caching. Build one small system end to end and instrument it.
• Weeks 5–8 — Evals: build the eval set, run error analysis, fix the top two failure categories, and document the delta. This is the portfolio centrepiece.
• Weeks 9–12 — Agentic depth and production: add a harness with fallbacks, guardrails against prompt injection, observability, and a cost/latency budget. Write it up publicly.

Three months of this produces more interview signal than a year of tutorials, because every step yields a number and a decision you can defend.

How to pick employers who are actually AI-first

Read the job description for evidence, not vocabulary. Good signs: they mention evals, production incidents, model cost, or a specific domain problem. Bad signs: a list of tool names and frameworks with no mention of what the system does or how it is measured.

In your own interview questions, ask: how do you evaluate model changes before shipping, who owns the eval set, and what was your last AI-related production incident? Teams that answer crisply have the craft. Teams that answer vaguely will expect you to build that discipline alone.

Roles we are hiring for right now are listed on our open roles board.

Frequently asked questions

Do I need a machine learning PhD to become an AI engineer? No. Ng's map puts building, fundamentals, agent use and product sense ahead of research depth. Machine learning foundations matter — bias/variance, error analysis, data engineering — but at working depth, not research depth.

What is the single strongest portfolio project? One system with a real eval set, a measured before-and-after improvement, and a written error analysis. It demonstrates the trait Ng calls the biggest differentiator, and almost no other applicant has it.

Is AI engineering experience required, or can a backend engineer transition? Backend engineers transition well, because software engineering fundamentals are the second skill on the map and the hardest to fake. Add evals and grounding on top of that judgment and you are a credible candidate within a quarter.

How should I handle AI tools during the interview? Use them if the role uses them, and narrate your verification. Interviewers are scoring supervision, not typing speed.