How do you become an AI engineer from software engineering?
Short answer
A software engineer becomes an AI engineer by adding skills in this order: calling models reliably, evaluating their outputs, grounding them with retrieval, enough classical machine learning to choose the right approach, and running AI features in production, with fine-tuning and agents last. Prove each skill with a small shipped project, not a certificate. Allow roughly six to nine months of part-time study; that is a worked estimate, not survey data.
By Dr. Dulan Dias, co-founder and CTO of Tellme AI
Key facts
- In Stack Overflow's 2026 Developer Survey, 73% of respondents said they use coding assistants and agents daily.
- In the same 2026 survey, 48% of respondents said they trust AI when they can easily validate its answers.
- LinkedIn's Jobs on the Rise 2026 report names large language models, retrieval-augmented generation and natural language processing as the most common skills among AI engineers in the UK.
- Rule #1 of Google's Rules of Machine Learning, by Martin Zinkevich, is “Don't be afraid to launch a product without machine learning.”
- The OWASP Top 10 for LLM Applications, 2025 edition, ranks prompt injection first, as LLM01:2025, and also lists unbounded consumption and excessive agency.
- NIST released its voluntary AI Risk Management Framework 1.0 on 26 January 2023 and a Generative AI Profile, NIST AI 600-1, on 26 July 2024.
- Chapter 18 of From Model to Product says machine learning is essential but no longer sufficient: the AI engineer also needs software engineering, system architecture, data engineering and infrastructure.
Definitions
- Golden set
- A fixed collection of real inputs with known good outputs or grading rules, used to score a model, prompt or retrieval change before it ships.
- Retrieval-augmented generation (RAG)
- Fetching relevant documents or records at request time and placing them in the model's context, so answers rest on your data rather than on what the model memorised.
- Embedding
- A vector of numbers that represents the meaning of a piece of text or an image, so similar items can be found by measuring the distance between vectors.
- Prompt injection
- An attack in which text from a user or a retrieved document overrides the instructions the application gave the model.
What carries over from software engineering?
Most of the job carries over. AI features are still software, and APIs, data models, tests, CI/CD, observability and incident handling decide whether they work in production. Chapter 18 of From Model to Product says it directly: machine learning remains essential but is no longer sufficient, and the engineer also needs software engineering, architecture, data engineering and infrastructure.
What is new is narrower than it looks. Outputs are probabilistic, so you test with example sets and scores instead of exact assertions, and quality depends on the context you assemble at request time. Every model call has a cost and a latency to budget. And the attack surface now includes any text your users and documents feed the model.
Which skills should you learn, and in what order?
Learn the skills in an order where each one makes the next measurable. This sequence suits someone who already ships software:
- Call models reliably. Use a hosted LLM API with structured output, timeouts, retries and streaming, and log tokens and cost per call from day one. The LLM API reliability guide covers the patterns.
- Evaluate before you optimise. Build a golden set from real inputs, score outputs with code checks, rubrics and, with care, a model as judge, and run it in CI. Every later decision depends on it; see how to evaluate an LLM application.
- Ground answers with retrieval. Learn embeddings, chunking, keyword and vector search, reranking and citations, and measure retrieval separately from answers. Start in Postgres; see pgvector vs a vector database.
- Learn enough classical machine learning to choose. Train and evaluate a small classifier, and understand train and test splits, precision and recall, overfitting, leakage and drift. That is how you know when a rule or a cheap model beats an LLM; see rules vs machine learning vs LLMs.
- Run it in production. Package and deploy a model service, monitor quality, latency and cost per request, add fallbacks, and defend against the OWASP LLM risks; see deploying a model with FastAPI and Docker.
- Adapt models last. Fine-tuning, open-weight hosting and agents come once evals show prompting and retrieval are not enough; RAG vs fine-tuning helps decide.
Which projects prove those skills?
Two or three projects done properly and written up honestly say more than a list of courses. Projects in a domain you already know make stronger evidence than generic chatbots.
| Project | Proves | Show in the write-up |
|---|---|---|
| A support-ticket triage service comparing a rules baseline, a small classifier and an LLM on the same labelled tickets | Approach selection and classical ML | Accuracy, latency and cost per ticket for each option, and which you would ship |
| A question-answering service over a public document set, with citations and a way to decline | Retrieval and grounding | Retrieval quality, answer quality on a golden set, and the questions it correctly refuses |
| An evaluation harness in CI that blocks a prompt change when the score drops | Evaluation | A before and after report from a real regression it caught |
| One of the above deployed behind an API with monitoring, a fallback and a per-user cost limit | Production engineering | The dashboard, a failure you simulated and the cost per thousand requests |
| A security pass against the OWASP LLM Top 10 | Security | The prompt injection attempts you tried and the controls that stopped them |
What can you skip, at least at first?
Skip anything that does not make your next project better. Google's first rule of machine learning is not to be afraid to launch without it, and the same discipline applies to LLM features.
- Training large models from scratch. Reading how it is done helps; doing it is research work applied roles do not need.
- Heavy maths up front. Learn the statistics behind evaluation as you need it; calculus can wait.
- Collecting frameworks. Orchestration libraries change quickly; learn one well and know what it does underneath.
- A dedicated vector database for a first project. Postgres with pgvector is enough to learn retrieval properly.
- Kubernetes. One container on a managed service is enough until traffic proves otherwise.
- Agents before evaluation. Multi-step agents multiply every failure mode; build them once you can measure single calls.
How long does it take to become an AI engineer?
No reliable survey measures how long the switch takes, so treat this as a worked estimate. Assume a working back-end or full-stack engineer, already comfortable with Python or TypeScript, HTTP APIs, SQL and Git, studying about 8 hours a week alongside the day job.
On those assumptions the stages below add up to 200 to 320 hours. At 8 hours a week that is 25 to 40 weeks, roughly six to nine months. Engineers who can practise on a real AI feature at work often move faster, because production teaches what tutorials skip; someone new to software should expect much longer.
| Stage | Estimated hours |
|---|---|
| Calling models reliably | 30 to 50 |
| Evaluation and testing | 40 to 60 |
| Retrieval and grounding | 40 to 60 |
| Classical machine learning fundamentals | 50 to 80 |
| Production: deploy, monitor, secure | 40 to 70 |
| Total | 200 to 320 |
How do you land the first AI engineering role?
The shortest route is usually inside your current job: volunteer for the AI feature, own its evaluation and its behaviour in production, and you have real experience before you change title. In the UK, LinkedIn's 2026 data puts most AI engineer jobs in London, Manchester and Cambridge, with LLMs, RAG and NLP the most common skills. Be ready to explain how you would measure quality, control cost and handle failure, not just which model you would pick.
From practice: at EchonLabs, Dr. Dulan Dias built Legify.ai, an AI legal-contract platform whose retrieval-augmented generation engine is organised by jurisdiction. The lesson for a portfolio is that retrieval design, here deciding which jurisdiction's documents may answer a question, is real engineering worth showing. For reading, the best AI engineering books are grouped by the stage you are at.
Frequently asked questions
Can you become an AI engineer without a machine learning degree?
Most applied AI engineering roles care more about shipped work than about a machine learning degree. A software engineer who can show a deployed AI feature with an evaluation set, monitoring and cost controls has the evidence those roles look for. Fill gaps in statistics and classical ML as you go, because interviews and real projects will test them.
Do you need a master's or PhD to become an AI engineer?
A master's or PhD is not needed for AI engineering roles that build products on existing models. Postgraduate degrees matter most for research roles that invent new models or training methods. A degree can still help if you want depth in a field such as healthcare or vision, but for applied roles a strong portfolio usually counts for more.
Should you learn PyTorch to become an AI engineer?
PyTorch is worth learning to a working level once the basics are in place, because fine-tuning, open-weight models and many research codebases use it. It is not the first step for a software engineer: calling hosted models, evaluation and retrieval come earlier. Learn it when a project needs a model you train or adapt yourself.
Is Python required for AI engineering?
Python is close to required for AI engineering, since most model libraries, evaluation frameworks and data tools are built for it. Strong TypeScript, Java or Go engineers can ship AI features in those languages through provider SDKs, but reading and modifying Python is expected in most teams. A few weeks of focused practice covers the gap for an experienced engineer.
Are AI engineering certifications worth it?
AI engineering certifications are worth little on their own. A cloud provider's AI certification can help you learn that platform's services and get past some recruiters' filters, but hiring managers judge AI engineers on what they have built and measured. Spend most of your time on projects with evaluation results and production details you can discuss.
Can a front-end developer become an AI engineer?
A front-end developer can become an AI engineer, but usually needs more back-end work first: APIs, databases, queues, deployment and observability, since that is where most AI engineering happens. Front-end skills remain useful for streaming responses, showing citations and designing how users recover from wrong answers. Budget extra months for the server side.
How do you get AI engineering experience if your company does not use AI?
Look for a small internal problem that AI could solve measurably, such as triaging tickets, searching documentation or summarising reports, and build it with an evaluation set and a cost estimate. Internal tools carry lower risk and still teach production lessons. If that is not possible, build a portfolio project in your company's domain using public data.
Sources
About the author
Dr. Dulan Dias is co-founder and Chief Technology Officer of Tellme AI in Belfast; founder of EchonLabs, grown from 2017 to more than 60 people before its acquisition in 2026; holder of a PhD in AI and data science from Queen’s University Belfast; and the author of From Model to Product.