roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Walk into AI Engineer interviews ready to design and defend production LLM systems. Spans Python and SQL, RAG and agents, evaluation and observability, and deployment.
7927 of 7927 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 7927 new. 0 cut, outside your plan.
Most AI engineer loops keep the outline of a software engineering loop and change what goes into each round. The coding round is in Python, sometimes an algorithm problem and increasingly a practical task: call a model API, validate its structured output, retry when it fails. A fundamentals conversation follows, then, from middle level up, a design round around a retrieval or agent system, and a behavioral conversation. The questions are specific. Why can a hosted model at temperature zero still answer the same prompt two ways, when the only thing that changed is the batch it was served in? Why do long prompts cost more to serve, in money and GPU memory? How would you chunk a document set, and show that a reranker earns the latency it adds? What stops a retrieved document from telling the model to call a tool? A junior answer shows the pieces working: an embedding, a similarity search, a prompt that cites its sources. A senior answer starts from measurement: a fixed evaluation set every prompt change must pass, traces that show which step produced a wrong answer, a cost per request, and a reasoned choice between better retrieval, fine-tuning and a larger model.
The plan starts with what you write code in and with: [Python](/topics/lang-python) down to generators and asyncio, [algorithms](/topics/found-dsa) for the coding round, and [Git](/topics/dev-git) with GitHub around both. Only then does it reach the model. [LLM methods](/topics/found-ai-llm-methods) lead, because prompting, embeddings, RAG and agents are the vocabulary the later modules speak, and they settle whether to [call a hosted model or run open weights](/topics/found-ai-engineering-model-sourcing-hosted-vs-open); [model hubs](/topics/ai-model-hubs) follow as the place those open weights come from. Retrieval needs a store, so [running a vector store](/topics/db-vector-concepts) comes next, how its index is sharded, replicated and rebuilt, then relational concepts and [SQL](/topics/lang-sql), where metadata, permissions and source records usually stay. With models and data in place the plan builds the application around them: tools exposed over [MCP](/topics/proto-mcp) and an API layer on HTTP, with [Server-Sent Events](/topics/proto-sse) carrying the token stream to a browser and FastAPI serving it. Then the question turns from whether it runs to whether it works. The [evaluation](/topics/found-ai-engineering-evaluation) and [tracing](/topics/found-ai-engineering-observability) taught among the LLM methods come well before testing and [application security](/topics/found-appsec), because output that varies cannot be tested until it can be scored. Deployment starts from cloud concepts and goes below the provider API, into [serving mechanics](/topics/ai-serving-concepts) such as continuous batching and quantization, then the inference servers built on them, [vLLM](/topics/ai-vllm) first, and Docker to package what you deploy. Working with coding assistants comes before the close: [system design](/topics/found-system-design) and resilience patterns for the senior round, then the [behavioral round](/topics/found-behavioral).
With a few weeks to go, narrow level to the seniority you are interviewing at, raise importance to must know, study by frequency, and keep daily time for the coding round, which an AI title does not remove. With a few months, study by topics and build one small retrieval application as you go: fundamentals questions are easier to answer about a pipeline you have run. Juniors and career changers should get Python, the LLM methods and one RAG pipeline solid before anything else. Seniors who already ship LLM features can move quickly through the opening modules and spend their hours on evaluation, serving and system design, where the design round expects numbers: tokens per request, a latency budget, the GPU memory a model needs. Fine-tuning appears here as one method among several; training models from data in depth, with loss functions, training loops and experiment tracking, belongs to the [Machine Learning Engineer roadmap](/roadmaps/machine-learning).
Most AI engineer loops keep the outline of a software engineering loop and change what goes into each round. The coding round is in Python, sometimes an algorithm problem and increasingly a practical task: call a model API, validate its structured output, retry when it fails. A fundamentals conversation follows, then, from middle level up, a design round around a retrieval or agent system, and a behavioral conversation.
The questions are specific. Why can a hosted model at temperature zero still answer the same prompt two ways, when the only thing that changed is the batch it was served in? Why do long prompts cost more to serve, in money and GPU memory? How would you chunk a document set, and show that a reranker earns the latency it adds? What stops a retrieved document from telling the model to call a tool?
A junior answer shows the pieces working: an embedding, a similarity search, a prompt that cites its sources. A senior answer starts from measurement: a fixed evaluation set every prompt change must pass, traces that show which step produced a wrong answer, a cost per request, and a reasoned choice between better retrieval, fine-tuning and a larger model.
The plan starts with what you write code in and with: Python down to generators and asyncio, algorithms for the coding round, and Git with GitHub around both. Only then does it reach the model. LLM methods lead, because prompting, embeddings, RAG and agents are the vocabulary the later modules speak, and they settle whether to call a hosted model or run open weights; model hubs follow as the place those open weights come from.
Retrieval needs a store, so running a vector store comes next, how its index is sharded, replicated and rebuilt, then relational concepts and SQL, where metadata, permissions and source records usually stay. With models and data in place the plan builds the application around them: tools exposed over MCP and an API layer on HTTP, with Server-Sent Events carrying the token stream to a browser and FastAPI serving it.
Then the question turns from whether it runs to whether it works. The evaluation and tracing taught among the LLM methods come well before testing and application security, because output that varies cannot be tested until it can be scored. Deployment starts from cloud concepts and goes below the provider API, into serving mechanics such as continuous batching and quantization, then the inference servers built on them, vLLM first, and Docker to package what you deploy. Working with coding assistants comes before the close: system design and resilience patterns for the senior round, then the behavioral round.
With a few weeks to go, narrow level to the seniority you are interviewing at, raise importance to must know, study by frequency, and keep daily time for the coding round, which an AI title does not remove. With a few months, study by topics and build one small retrieval application as you go: fundamentals questions are easier to answer about a pipeline you have run.
Juniors and career changers should get Python, the LLM methods and one RAG pipeline solid before anything else. Seniors who already ship LLM features can move quickly through the opening modules and spend their hours on evaluation, serving and system design, where the design round expects numbers: tokens per request, a latency budget, the GPU memory a model needs.
Fine-tuning appears here as one method among several; training models from data in depth, with loss functions, training loops and experiment tracking, belongs to the Machine Learning Engineer roadmap.