Switch into AI and ML from a regular software job

Pick a destination title

These are different jobs:

• ML / AI engineer — ship models and pipelines (training, eval, inference). Closest jump from SWE.
• Applied scientist — more experimentation and papers-adjacent work.
• ML platform / MLOps — serving, GPUs, Kubernetes, observability. Strong fit if you already do backend and infra.
• Researcher — frontier models, publications. Harder without a research track record.

Target the first or third unless you already have papers.

Skills that hiring managers actually test

• Python plus PyTorch (or a stack you can defend)
• Train a small model end to end: data → train → eval → a simple API
• LLM application patterns: prompting, RAG, eval sets, not only ChatGPT screenshots
• Basics of latency, cost, and failure: what happens when the model is wrong in production
• SQL and a cloud (Azure is the obvious Microsoft-shaped choice; AWS/GCP count)

A project that counts

One complete project beats five tutorials. Example: a classifier or RAG tool on a real dataset, with an eval sheet (precision/recall or a labelled LLM rubric), a README, and a deploy. Put the GitHub link on the resume. If your current team will let you add an eval harness or a model-backed feature, that is stronger than any side project.

How to read AI JDs

If required says “6+ years technical engineering” plus ML in preferred, that is a software seat with ML on the job. Apply. If required is a PhD and a NeurIPS/ICML track record, that is research; don’t force it. “Model optimization,” “distributed training,” and “LLM eval” in the responsibilities means they want systems plus ML, not a prompt-only portfolio.

First 90 days after you land it

Ask how they evaluate models, where the training data lives, and what “good” means for the product. Ship a small improvement to eval or serving before you propose a new architecture. Credibility in AI teams still comes from production, not from slide decks about agents.