Roadmap
AI Engineer
An AI engineer is a software engineer whose hardest dependency happens to be a model. This route assumes you will call a model rather than train one, so it builds Python and HTTP first, then covers what the model is, how to call it from code, how retrieval and vector search work, and what an agent loop really is. Every stage is about the engineering around the model, because that is where the work turns out to be.
6 stages · 28 steps
Who this is for
Developers moving into applied AI, and beginners who want to build with language models but would rather understand the pieces than wire together a demo they cannot debug.
Python and HTTP first
Write Python that handles messy input and calls an external service without falling over.
- Language guidePythonAlmost every model provider's SDK and example is Python first.
- LessonPython FunctionsThe unit you will wrap every model call in, so retries and logging live in one place.
- LessonPython DictionariesA model request and its response are both nested key-value data.
- Concept guideError HandlingModel calls time out, rate-limit and return nonsense; this decides what your code does then.
- PlaygroundPython playgroundFor shaping a request or picking apart a response without writing a whole program.
- Concept guideHTTP and HTTPSEvery model call is an HTTP request, and most debugging happens at that level.
What the model is, before you call it
Describe in plain language what a language model does, so you can tell a model problem from a code problem.
- AI courseAI FundamentalsTokens, context windows and training, without the marketing.
- Concept guideMachine Learning FundamentalsThe training process behind the endpoint you are about to depend on.
- Concept guidePython for AI and Machine LearningThe array and dataframe layer you will meet as soon as you handle embeddings.
Calling a model from code
Make a model call from a program, handle its failures, keep the key safe, and get usable output on purpose.
- Concept guideWorking with LLM APIsMessages, parameters, streaming, cost and the failure modes to plan for.
- Concept guideREST APIsYou will consume one and then expose one; both halves use the same conventions.
- AI coursePrompt Engineering for Real WorkPrompting is part of the interface you are building, not a trick.
- Developer toolJSON Formatter & ValidatorFor checking the structured output a model claimed to return.
- Concept guideSecrets and Credential ManagementAn API key that bills per call does not belong in your repository.
Retrieval
Give a model the right context: embed text, store the vectors, search them, and assemble a grounded answer.
- Concept guideEmbeddings, ExplainedStart here; vector search and retrieval both rest on this single idea.
- Concept guideVector DatabasesWhere large numbers of those vectors get stored and searched, and what that costs.
- Concept guideRetrieval-Augmented Generation (RAG)The pattern that turns a model into something that answers questions about your data.
- Concept guidePostgreSQLMany teams start with the relational database they already run rather than a new one.
- Concept guideCaching StrategiesRepeated questions are the cheapest latency and cost win you will find.
Agents and the systems around them
Explain an agent loop honestly, and build the plumbing that keeps one from running away with your bill.
- Concept guideBuilding AI AgentsA loop, a set of tools and a stopping condition — read it after the API guide.
- AI courseAgentic AIThe same territory from the product side, including where agents are the wrong answer.
- Concept guideMessage QueuesLong model work does not belong inside a web request.
- Concept guideWebhooksHow a long-running job tells the rest of your system it finished.
- Concept guideDocker and ContainersTool-running code is much safer inside a container boundary.
Interview rounds for this role
Answer AI-specific questions and the ordinary engineering ones in the same interview.
- Interview prepAI & Machine Learning interview questionsStart with the round that is specific to this role.
- Interview prepBackend & API interview questionsThe job is backend engineering with a model attached, and the interview reflects that.
- Interview prepSystem Design interview questionsExpect to design a retrieval pipeline and be asked what it costs.
- Interview prepCoding & DSA interview questionsStill asked: a model in the stack does not remove the algorithm round.
Projects to build along the way
Briefs in the languages this roadmap uses.
- Number guessing gameHide a number, read guesses one line at a time, and say higher or lower until it is found.
- Gradebook reportTurn a list of students and scores into an aligned table with averages and letter grades.
- Text adventure driven by dataA small explorable world where the rooms live in a data structure, not in a chain of if statements.
- REST API for a notes appA small HTTP API with a real database behind it, honest status codes, and tests that hit every route.
- Mini search engineAn inverted index over your own documents, with ranking you can justify term by term.
Interview topics for this role
What you will need to learn elsewhere
This site cannot teach everything this role involves. Plan for these too:
- One model provider's own documentation and SDK. Model names, limits and prices change faster than any guide can track.
- Evaluation: a test set of real questions, a way to score answers, and evidence that a change actually improved them.
- Cost and latency control in production — token accounting, batching, timeouts, and a budget alarm.
- Running a vector store for real, or the vector extension of a database you already operate.
- Deploying and monitoring the service, with logging good enough to trace one bad answer back to its inputs.
- Privacy and policy: what data your employer permits you to send to a third-party model, and what you must log or redact.
- An orchestration framework's own tutorial if your team uses one; this layer changes shape often.
- Building something people use, then watching how they break it. Prompt injection is found in the wild, not in a guide.