TL;DR
For most AI engineering roles a portfolio of shipped projects beats a degree. The job is building production software with existing LLMs at the core (support bots, internal search, agents), not researching or training models. A realistic self-taught path runs about twelve months, and it only works if you ship something every phase.
The distinction that saves months
- ML researcher: invents and trains new models. Benefits from advanced degrees and heavy math. Small slice of the market.
- AI engineer: builds useful things with models that already exist. Rewards software skill, product sense, and shipping discipline. This is the role you can enter without a degree. You do not need mastery of linear algebra or calculus to start.
Actionable items (12-month path)
- Months 1-3, code properly. Get genuinely fluent in Python (variables, control flow, functions, files, calling APIs, error handling, reading others’ code). Learn Git and push to a public GitHub from day one. Build five tiny programs: calculator, file organizer, an API-calling script, a data cleaner, a CLI note taker. Write code every day.
- Months 3-5, master the LLM API. Work through the API, not the chat UI. Handle streaming, conversation history, output format control, rate limits, and errors. Learn the gap between “usually right” and repeatable production output. Learn tool use (function calling): give a model one function and make it call it correctly. Build a CLI text tool and a chatbot with memory.
- Months 5-7, build RAG. Chunk documents, create embeddings, store in a vector database, retrieve the most relevant chunks, feed them to the model for grounded, sourced answers. Build one end-to-end RAG app over real documents, add retrieval evaluation, and deploy it so a stranger can use it. This is portfolio project one.
- Months 7-9, build agents. An agent is a model in a loop with a goal and multiple tools that decides its next step from results. The hard, market-starved skill is reliability: failure handling, clear tool design, evaluation. Build a single-agent multi-tool system, then a small multi-agent system with explicit failure handling. Portfolio project two.
- Months 9-11, evaluation and deployment. Build evals that measure whether the system is actually good and whether a change helped (accuracy, relevance, consistency; model-graded or human review). Deploy properly: hosting, monitoring, load, cost tracking, catching failures before users (MLOps basics). Portfolio project three: a deployed system with evaluation and monitoring.
- Months 11-12, get hired. Write each project as a case study (problem, approach, what you measured, what you’d change). Build in public and post breakdowns. Apply broadly, including AI-augmented software roles as a realistic first step. Entry salaries commonly run from roughly 200K with experience, depending on company and location.
Key quotes
A demo agent is easy and a reliable agent is hard. The gap between the two is failure handling, clear tool design, and evaluation.
Reading about AI engineering is not becoming an AI engineer. The people who come out employed are the ones who shipped something every single phase.
You’re not learning to compete with the tools. You’re learning to command them.