Auto Apply to hundreds of jobs in a few hours

Español

日本語

AI engineer skills: what to learn first, and what can wait

Model names change every few months, so the skill that lasts is building software around a model you don't control. Learn to call a model, feed it the right context and measure what comes back. Fine-tuning and fancy agent frameworks can wait until someone is paying you to need them.

4,114 open jobs

Step one

Gets you the interview

Production Python with typed APIsMost AI engineer work is a FastAPI service with Pydantic models, async calls and retries around a model endpoint. Screeners want to see code that handles a timeout, not a notebook that worked once.

Calling model APIs with structured outputYou should know how to send a system prompt, stream tokens back, force a JSON schema and handle a refusal or a malformed reply. Take-home tasks nearly always start here.

Retrieval with embeddings and a vector storeChunking documents, choosing an embedding model and querying pgvector, Qdrant or Pinecone is the backbone of a lot of company chatbots. Knowing why hybrid search with keyword matching beats pure vector search on part numbers and names gets you noticed.

Reading and cleaning messy source dataPDFs with tables, scraped HTML and ticket exports full of signatures wreck retrieval quality. Showing you can parse and clean them tells a hiring manager you've built the real thing.

Step two

Gets you the offer

Building an evaluation setAnyone can demo a chatbot. The candidate who brings graded examples, an LLM-as-judge rubric they've checked against human labels and a regression run in CI is the one who gets the offer.

Tool calling and agent loopsLetting a model call your search, database or ticketing functions means designing tool schemas, capping the number of steps and deciding what happens when it calls the wrong one. Interviewers will ask you to sketch this on a whiteboard.

Tracing and debugging model callsWhen a user says the answer was wrong, you pull the trace in Langfuse, LangSmith or your own logs and see which retrieved chunk misled the model. Walking through a real trace in an interview beats any buzzword.

Guardrails and prompt injection defenseA support bot that leaks another customer's data or follows instructions hidden in a web page is a security incident. Knowing input filtering, output checks and least-privilege tool access shows you can ship something legal will sign off on.

Step three

Gets you promoted

Cost and latency engineeringCaching prompts, routing easy questions to a smaller model, batching and trimming context are what keep a feature alive once finance sees the bill. The engineer who cuts cost without hurting eval scores becomes the one leadership asks first.

Serving open-weight modelsRunning Llama or Mistral models on vLLM or Hugging Face TGI, with quantization and GPU sizing, matters when data can't leave your cloud. It moves you from wiring APIs to owning the platform.

Fine-tuning with LoRA, and knowing when not toParameter-efficient fine-tuning with PEFT fixes tone and format problems prompts can't. Senior engineers are paid to say when a better retrieval pipeline would do the same job for less.

Setting quality bars with productDeciding what error rate is acceptable for a medical summary versus a marketing draft is a product call with engineering evidence behind it. Leading that conversation is how you become the staff engineer or AI lead.

Certificates worth your time

Certificate	Best for	Effort	Worth it?
Microsoft Certified: Azure AI Engineer Associate	Engineers at companies that run on Azure and use Azure OpenAI or AI Search	a few weekends if you already build on Azure	The closest match to this job title of any certificate. Worth it on an Azure shop, ignored almost everywhere else.
Google Cloud Professional Machine Learning Engineer	Engineers building on Vertex AI who want proof of cloud ML depth	several weeks of study around a full-time job	Respected, but it leans toward classic model training. Take it if your team lives on Google Cloud.
AWS Certified AI Practitioner	Software engineers new to AI who work in an AWS company and want a first credential	a couple of weekends	Entry level and quick. It helps you get past a keyword filter, but it won't convince a technical interviewer on its own.

Exam content and names change often in this field, so check the current outline directly with Microsoft, Google Cloud or AWS before you book an exam.

Put it on your résumé like this

Weak

Experienced with LLMs, RAG and prompt engineering.

Strong

Added hybrid BM25 plus vector search and a 400-case eval suite to an internal policy assistant, lifting grounded-answer rate from 72% to 91% and cutting cost per query by 38% with prompt caching.

Questions people ask

Do I need to learn the math behind transformers?

Enough to explain attention, tokens and context windows in plain words, yes. You won't derive gradients on the job. Interviewers care more that you can say why a long context gets slow and expensive than that you can write the equations.

Should I learn LangChain or LlamaIndex first?

Learn to build a small retrieval app with plain API calls and a vector store first. Then pick up one framework so you can read other people's code. Plenty of teams rip frameworks out once the product matures, and interviewers like hearing that you know what's under the hood.

Is prompt engineering a real skill or a fad?

It's real, but it's a small slice of the job. Writing clear instructions and examples matters. What gets you hired is proving a prompt change helped, with an eval run, instead of eyeballing a few answers.

How do I practise without a big cloud budget?

Run a small open model locally with Ollama, use free API tiers for experiments, and store embeddings in Postgres with pgvector. Build one tool that answers questions over documents you know well, then write the eval set before you tune anything.

Got step one? Start applying. HeroApply matches you to roles that fit.

Start your trial