LinkedIn's Jobs on the Rise 2026 list puts AI engineer at number one in the US, and the same report says most people in the role moved over from software engineering or data science. That second part matters more to you than the ranking. The door is open to backend and fullstack developers, and most of the people walking through it did not start with a PhD.
I review a lot of applications for our AI roles. The pattern holds. Candidates who get interviews have shipped one or two real LLM features and can talk about what broke. Candidates who don't have a list of certificates and a chatbot demo that calls an API once.
Key Takeaways
- AI engineering at a US startup is mostly product engineering on top of foundation models: retrieval, tool calls, evals, cost and latency.
- Your backend skills carry over. APIs, queues, databases and logging are half the job.
- Evals separate hires from tourists. Very few candidates can show how they measured quality.
- If you already ship production code, 90 days of focused building is enough to become interviewable.
What the job actually is
Most startups are not hiring you to train models from scratch. The AI engineer at a Series A company takes a model from Anthropic, OpenAI or Google, wires it into the product, and makes it behave. LinkedIn lists the most common skills for the role as LangChain, retrieval-augmented generation (RAG) and PyTorch. On our own role pages for AI engineer and agentic AI engineer, the day-to-day looks like this:
- Design retrieval: chunking, embeddings, hybrid search, re-ranking.
- Build agents that call tools and recover when a tool fails.
- Write evals so a prompt change doesn't quietly break last week's cases.
- Watch cost per request and latency, because the CFO will.
Look at how much of that list is plain backend work. A Python or TypeScript developer who has run a production API already understands retries, timeouts, idempotency and queues. Those are exactly the parts most bootcamp graduates are missing, and they're the parts that break at 2 a.m.
The 90-day plan
| Days | Build | What you should be able to explain |
|---|---|---|
| 1-30 | A RAG app over a real document set (your country's customs rules, a public API's docs, a company handbook) with a small web UI | Why you chunked the way you did, what retrieval misses, what one query costs |
| 31-60 | An eval suite: 50 to 100 test questions with expected answers, scored automatically and spot-checked by hand | How accuracy moved when you changed chunk size, model or prompt |
| 61-90 | An agent: two or three tools (search, a database query, drafting an email) with failure handling, tracing and a spending cap | What happens when a tool times out, and how you know from the logs |
Pick one language and stay there. Python has the widest library support. TypeScript is fine if you're a React and Node developer aiming at fullstack AI roles. Switching halfway costs you a month.
Deploy it. A URL a reviewer can open beats a notebook every time. Then write a README that states the eval score at day 45 and at day 90. That one number puts you ahead of most applicants, because almost nobody measures.
If you want structured learning next to the building, our list of AI courses for software developers covers free and cheap options. Treat courses as support. The project is the credential.
A Concrete Version
This is a composite, built from the kind of profile we see often. Call her Daniela: a Python backend developer in Medellin, six years in, mostly Django and Postgres at a logistics company.
Days 1 to 30, she builds a question-answering tool over about 400 pages of public customs regulations. First version: fixed 1,000-token chunks, one embedding model, top-5 retrieval. On her own spot checks it answers confidently and wrong roughly a third of the time.
Days 31 to 60, she writes 80 test questions. Baseline: 58% correct. Section-aware chunking lifts it to 71%. A re-ranking step gets it to 79% and adds around 400 ms per query. She writes that tradeoff into the README instead of hiding it.
Days 61 to 90, she adds two tools: a tariff lookup against a small Postgres table and a "draft a summary email" action. Every tool call is logged, spend per session is capped, and a timeout returns a partial answer with a warning.
In her interviews nobody asks about courses. They ask whether re-ranking was worth 400 ms. She has an answer with numbers, and that is most of what the interviewer wanted to find out.
The Honest Counterpoint
This plan assumes you already ship production code. With less than two or three years of real backend work, the faster route to an AI engineer title usually runs through a regular backend job first. LinkedIn's data points the same way: the report puts median prior experience for AI engineers at 3.7 years. Startups hiring remotely from LATAM want people who can own a feature alone, and a junior developer with a RAG demo rarely clears that bar.
It also won't make you an ML researcher. Training models and working on model internals takes deeper math and usually a different path, like the one in our machine learning course guide. And the tooling moves fast. The framework you learn in month one may be out of fashion by month six. Retrieval, evals and failure handling don't go stale.
Frequently Asked Questions
Do I need a master's degree to become an AI engineer?
Not for the product-focused roles most US startups hire for. What gets you hired is shipped work and the ability to explain tradeoffs. Research roles are a different story.
Python or TypeScript for AI engineering?
Python if you're undecided, since most eval and data tooling starts there. TypeScript works well if you're already a strong React and Node engineer going after fullstack AI roles.
How good does my English need to be?
Good enough to defend a design decision on a call and in writing. Our job posts ask for B2 and up, and client-facing roles lean closer to C1. AI engineers write a lot of design notes, so written English counts as much as spoken.
The Bottom Line
Pick one real dataset, build retrieval, measure it, turn it into an agent, and write down the numbers. Ninety days of that makes you interviewable for roles many LATAM developers assume are out of reach. Once the project is live, apply to our AI engineer role or browse every open role.
---
Roberto Espinoza is CEO of Ruzora, which places senior LATAM engineers with US startups. Browse open roles.
