Start here: experienced engineer

Keep your software engineering strengths. You do not need to repeat every introductory programming course. The new work is understanding model uncertainty, testing behavior that is not perfectly repeatable, and managing the quality/cost/latency tradeoff.

Your first week · 10 focused hours

1. 1 hour: choose a role and run the foundations diagnostic. If you can already build, test and deploy an API, skip the beginner sequence.
2. 2 hours: watch LLM Foundations; write a short explanation of tokens, context, inference and embeddings. Watch the lecture ↗.
3. 3 hours: build one structured extraction endpoint with a current provider SDK. Validate the output against a schema; record model and prompt versions.
4. 2 hours: write 10 fixed examples, including malformed input and a case that should return “unknown”. Compare with a simple rule-based baseline.
5. 2 hours: add timeout handling, a fake provider for tests and a README with observed failures.

What to reuse and what to learn

• Reuse your preferred web stack. Python is the default here because the exercises and AI ecosystem are accessible; a strong TypeScript engineer can ship the main project in TypeScript and build Python reading fluency.
• Treat a prompt, dataset, retrieval configuration and model choice as versioned inputs to a system. Measure changes instead of relying on a single impressive answer.
• Build one retrieval application and one bounded tool workflow. Learn a framework only when explicit orchestration becomes hard to manage.
• Planning allowance: 120–180 focused hours including a capstone and interview preparation, about 12–18 weeks at 10 hours/week. Prior domain experience may shorten discovery; difficult debugging may lengthen implementation.

A useful stretch assignment

Take an existing feature and add an AI suggestion with an explicit human decision point. Compare end-to-end task success with the original workflow. Keep authentication, data ownership and failure recovery under ordinary application control.