A practical ML learning path for software engineers

If you already ship web apps, you have a head start: data pipelines, debugging, and API design transfer directly to ML systems. The gap is mostly math intuition and experiment discipline.

Phase 1 — foundations: refresh linear algebra (vectors, matrices, dot products), basic probability, and Python numerics with NumPy. Pair this with Andrew Ng's ML fundamentals or fast.ai's practical first pass.

Phase 2 — deep learning: learn PyTorch tensors, autograd, and a simple training loop. Build a tiny classifier on a dataset you care about — not MNIST unless you must. Log everything.

Phase 3 — applied AI: Hugging Face for models, LangChain or raw APIs for agents, vector DBs for retrieval. Focus on evaluation: if you can't measure quality, you can't improve it.

Ship one end-to-end project: ingest → embed → retrieve → respond, with tests and a README. That artifact teaches more than a dozen half-finished notebooks.