The AI Engineer Career Kit
A roadmap you can fail, 12 interview answers written out in full, and the templates that turn work you have already done into something an interviewer believes.
The honest limit: nothing here gets you a job. It shortens the distance between the work you have done and an interviewer believing you. If you have not built anything yet, start with the roadmap below — the first two stages are free — and come back.
The roadmap
Six stages. Each one has a goal, the thing that goes wrong at that stage, and — the part that matters — a set of checks you either pass or do not.
The Python you actually need
Write Python that another person can run without you sitting next to them.
The trap
Learning Python from tutorials indefinitely because each one is comfortable. The cure is a script that has to survive a real API returning a 500 halfway through — that single experience teaches more than three more courses.
What to learn
- Types, collections, comprehensions, and when a comprehension stops being readable
- Functions, arguments, defaults, and why a mutable default argument bites
- Files, JSON, and HTTP requests with error handling that is not bare except
- Virtual environments and a requirements file that pins versions
- Enough of the standard library to stop reaching for a package: pathlib, datetime, collections, itertools
Done when — all of these are true
- You can write a script that reads a CSV, calls an API for each row, handles a failure on row 40 without losing rows 1 to 39, and writes results to a file.
- Somebody else can clone your repository, follow the README, and run it first time.
- You can explain what your virtual environment is for without saying "it just works".
Your first LLM application
Ship something that calls a model and is useful to one real person who is not you.
The trap
Starting with a framework. It works on day one and leaves you unable to debug on day thirty, because you never learned what it was doing. Call the API raw first — it is fifteen lines — then adopt a framework knowing what it replaced.
What to learn
- Calling a model API directly, without a framework, at least once
- Prompt structure: instructions separated from data, and why that separation is a security boundary and not a style preference
- Structured output — asking for JSON and validating it rather than hoping
- Token costs and latency: what your call actually costs and how long it takes
- Handling the failure cases: rate limits, timeouts, and a response that is not the JSON you asked for
Done when — all of these are true
- You have called a model without a framework and can explain what LangChain would have added.
- Your application handles a malformed model response without crashing, and you have seen it happen.
- You can state the cost per request in rupees or cents, and how you measured it.
- One person who is not you has used it and told you something you did not expect.
Build next: pdf qa agent
Retrieval that actually works
Build a RAG system and be able to prove whether it is any good.
The trap
Judging the system by asking it questions and reading the answers. It feels like testing and is not — you unconsciously ask questions you know it handles. Thirty fixed questions with known answers, measured before and after every change, is the whole difference between engineering and vibes.
Build next: pdf qa agent, rag evaluation harness, semantic search over code
Agents, tools, and knowing when not to
Build something that uses tools, and develop the judgement to argue against doing so.
The trap
Building an agent for a problem with a known sequence of steps. If you can write the steps down, write the steps down — an agent adds cost, latency and non-determinism to buy flexibility you are not using. Interviewers ask this precisely because so few candidates can answer it.
Build next: research agent with tools, support ticket triage, mcp server for your docs
Getting it off your laptop
Deploy something, watch it break in a way you did not predict, and fix it.
The trap
Treating deployment as the last step. It is the step that teaches you what your architecture got wrong, and doing it last means learning that when it is expensive to change. Deploy the ugly version in week one.
Build next: llm cost observability, invoice extraction pipeline
Being believed
Turn work you have done into an account an interviewer trusts.
The trap
Rehearsing what went well. Everyone does that and it is unfalsifiable, so it carries no information. The candidate who says "the first version chunked at 1000 characters and recall was 60%, so I chunked on clause boundaries and it went to 85%" has proved they were there. Nobody invents that detail.
12 interview questions, answered
Written as answers rather than as bullet points about what to mention. Read one aloud once and you learn how long an answer should be — a candidate holding a list of keywords never finds that out, and it is audible.
Templates
Already bought it?
Enter the email you paid with and we will send an access link.