TL;DR: You become an AI engineer by building the systems AI engineers are hired to build, in the order employers care about: a feature on top of a hosted model, a retrieval pipeline with evaluation, a model served on a GPU within a budget, then fine-tuning and agents. Most people can get there from a software engineering job in about 90 days of focused evenings. Interviews test the artifacts and the reasoning behind them far more than credentials, so the plan below is organized around what you will have built by the end of each month.
Consider a backend developer who has spent six years shipping web services. She has never trained a model, and she has watched her company bolt a chat assistant onto its product using an API and a prompt. The assistant hallucinates, the bill is unpredictable, and nobody on the team can say why it fails on Tuesdays. Her manager asks whether she could own it.
That request is how most AI engineering careers begin. The role is less about inventing models and more about making them dependable inside a product, and the people who fill it are overwhelmingly engineers who crossed over rather than researchers who came down. This guide is written for that crossing: what the job is, what to learn and in what order, what to build so the learning leaves evidence, and how hiring actually works.
Start here
This guide covers the career side. For the full technical sequence stage by stage, read the AI engineer roadmap. For what the job pays and which skills move the number, read AI engineer skills and salary.
What an AI engineer does all day
Strip away the titles and the work is five recurring activities. Choosing a model for a task and defending the choice with numbers. Getting the model to behave, through prompting, structured output, and, when that runs out, fine-tuning. Building the system around the model: retrieval over the company's documents, tools the model can call, memory, guardrails. Serving it on GPUs so latency and cost stay inside a budget. And evaluating it, so the team knows when a change made things worse before customers do.
None of those activities is research. All of them are engineering with a probabilistic component in the middle, which is the part that feels alien on day one and routine by month three. If you can already ship a web service with tests, monitoring, and a cost line, you have most of the muscle; what you are adding is a new kind of component and the judgment to work with it.
The three routes in, and what each one is missing
From software engineering. The most common route and the shortest. You already have system design, testing, deployment, and cost instincts. What is missing is the model: how it works internally, why it fails, and how to evaluate a component whose output varies. Your first month should be almost entirely stage 1 of the roadmap (build a transformer from scratch) plus API engineering, and you will be surprised how quickly the rest falls into place.
From data science or ML engineering. You know models, training loops, and evaluation. What is missing is production engineering: serving, containers, CI, observability, and the system-level patterns (retrieval, tool use, agents) that dominate LLM work. Your gap is stages 3, 6, and 7, and you should start with serving because it is the skill hiring managers find hardest to source.
From a student or fresher background. You have time and no artifacts. The route that works is to skip breadth and go straight to building the five artifacts in the plan below, because a fresher with a served model and a RAG pipeline with evaluation numbers is a stronger candidate than a graduate with three certificates. This route is especially relevant in India, where the volume of "AI engineer" applicants is high and the number who can show a deployed system is small.
The skills that get you hired, in priority order
- Python and one ML framework, at a working level. PyTorch is the safe choice. You need to read model code, write a training loop, and debug shapes.
- How language models work. Tokenization, attention, the decoder-only transformer, sampling. Deep enough to predict behavior.
- API engineering around models. Prompt structure, few-shot examples, structured output (constraining the model to return JSON your code can validate), tool calling, token counting, caching, and defending against prompt injection.
- Retrieval-augmented generation. Embeddings, hybrid search (dense vectors plus keyword search), reranking, grounded answers, and evaluating retrieval separately from generation.
- Serving and inference economics. Quantization, batching, vLLM, and the ability to state tokens per second and cost per request for a system you run.
- Evaluation and experiment tracking. Building an eval set, LLM-as-judge with its biases, MLflow, and regression gates in CI.
- Fine-tuning. LoRA and QLoRA on one GPU, evaluated against the base model.
- Agents. ReAct loops, MCP tool servers, orchestration, memory, guardrails.
The first five are what junior and mid-level job descriptions ask for. Six through eight are what separates senior candidates, and they are the ones most applicants cannot demonstrate.
The 90-day plan
The plan assumes eight to ten hours a week, which is what a working developer can sustain. Each month ends in artifacts, and the artifacts are chosen because they are what interviewers ask about.
Month one: the model and the API. Week one and two, build a decoder-only transformer from scratch and train it on a small corpus until it generates coherent text; you will reference this build in every interview for years. Week three and four, ship a feature on a hosted model with a cost ceiling, a retry and timeout policy, structured output validation, and a prompt injection test. Artifacts: the transformer repository with a training curve, and the API feature with a cost dashboard.
Month two: retrieval and serving. Weeks five and six, build a retrieval pipeline over a real document set with hybrid search and a cross-encoder reranker, plus a small evaluation set that scores retrieval and answers separately. Weeks seven and eight, quantize an open model, stand up vLLM, and record throughput, latency, and time to first token at several batch sizes. Artifacts: the RAG system with its eval numbers, and the serving benchmark with a chart.
Month two and three on real GPUs
Build hybrid retrieval with cross-encoder reranking, serve a model on vLLM and measure it, then fine-tune Llama 3 8B with LoRA and compare it against the base model, each in a browser-based GPU environment with the numbers you will quote in interviews.
Month three: fine-tuning, tracking, and an agent. Weeks nine and ten, fine-tune an 8B model with LoRA on your own instruction data, log every run in MLflow, and evaluate against the base. Weeks eleven and twelve, build a ReAct agent with a tool-call budget, expose your tools through an MCP server, and score the agent's runs with an LLM-as-judge harness. Artifacts: the adapter weights with the comparison table, the MLflow project, and the agent with its evaluation report.
Five artifacts, each with numbers attached. That is a stronger portfolio than most people with the title currently have.
What each artifact proves in an interview
| Artifact | The question it answers | Typical follow-up |
|---|---|---|
| Transformer from scratch | Do you understand what the model is doing? | Why does greedy decoding loop, and what does temperature change? |
| API feature with cost ceiling | Can you ship a safe, budgeted LLM feature? | How did you handle an injection in a retrieved document? |
| RAG pipeline with eval | Can you build the pattern most teams need? | Your retrieval recall dropped after a reindex; what do you check first? |
| vLLM serving benchmark | Can you run models within a budget? | What did continuous batching change in your numbers? |
| LoRA fine-tune plus MLflow | Can you improve a model and prove it? | When would you keep the adapter separate instead of merging? |
| Agent with eval harness | Can you build something that acts, and measure it? | How do you stop a tool loop from running forever? |
How the interviews actually work
Most AI engineer loops have four parts, and only one of them looks like a traditional coding interview.
The portfolio walk. You are asked to explain one of your systems end to end, and the interviewer pushes on the decisions: why this chunk size, why a reranker, why that quantization level, what the eval set looked like. This is where the artifacts pay off. Candidates who built the thing answer with specifics; candidates who followed a tutorial stall at the second question.
The system design round. A prompt like "design an assistant over our support documentation for 10,000 users." They want to hear retrieval architecture, evaluation, cost and latency budgets, failure handling, and what you would not build in version one. Practice by designing the systems you built as if for ten times the scale.
The debugging round. A model behaves badly and you diagnose it. Common scenarios: answers are fluent and wrong (retrieval or grounding), latency spikes under load (batching, KV cache, or context length), costs doubled (context growth, retries, or a caching regression), the agent loops (tool budget, stop conditions).
The coding round. Usually Python, often on data manipulation, tokenization, or a small piece of a pipeline. Rarely the algorithmic puzzle of a general software loop.
Certifications come up in one place: the resume screen. They help you reach the loop more often, especially at larger companies and for roles that name a vendor stack, and they do not carry you through it. The certification guide covers which ones are worth the fee in 2026.
Mistakes that stall the transition
Learning breadth before building depth. Survey courses feel productive and leave nothing behind. One built system teaches more than five completed courses and gives you something to talk about.
Skipping how the model works. Engineers who treat the model as a black box spend their first year debugging by superstition. The transformer build is a weekend; the superstition is a career.
Treating evaluation as optional. Every artifact in the plan has numbers attached because interviewers ask for them and because production requires them. A RAG pipeline without an eval set is a demo.
Never touching a GPU. Serving is the scarcest skill in the stack, and it cannot be learned from a diagram. Rent or borrow a GPU for month two if your laptop cannot run one.
Chasing the newest framework. Frameworks rotate every six months. The patterns (retrieval, tool use, evaluation, serving) have not changed in three years. Learn the patterns and treat frameworks as interchangeable.
Frequently asked questions
Key takeaways
0/6 completedNext steps
Pick your route, then start the 90-day plan with the transformer build. If you want the same sequence with graded work and real GPUs, the AI Engineer path runs it as 8 modules with 26 GPU labs and 6 graded projects; the curriculum is open to browse, and the lessons, labs, and projects are included with Preporato Pro. For the detailed technical sequence, keep the roadmap open alongside this guide.
Sources:
- Built In: AI Engineer salary in US, 2026
- Glassdoor India: AI Engineer salaries
- NVIDIA Certified Associate: Generative AI LLMs