Questions.

A place to practise AI engineering the way it is done: you build and fix model calls, retrieval, RAG, tool calling, agent loops, MCP servers and evals, and your work is graded against real runs. It is made by Deep-ML.

Graduates and working software engineers becoming AI engineers, and anyone preparing for AI engineering interviews. You should be comfortable writing Python.

Start the curriculum: one path through every open section, in order, with a button that takes you to the next problem you are up to. If you would rather browse, the catalog filters by level, format and difficulty, and offers today's problem.

No account is needed to read problems or run code. You need one to submit project steps, to keep your progress across devices, to save your agent to your account, and to push your work to GitHub.

No. Problems that call a model run on Gemma 4, which we run and pay for; your monthly credits cover it. You never paste a key into the site.

Five formats. Function: implement or debug one piece of an agent, graded by tests. Agent loop: close a loop, graded turn by turn against a recorded trace. Repair session: fix a broken agent in a small repo, with the coding agent you built. Tool design: rewrite tool descriptions until every request routes right on every one of three runs. Error analysis: label failed runs, find the first bad turn, and pick the fix.

Most problems are graded against a real model, because that is what you ship against. Some run against a scripted model so that your code is the only thing that varies; the catalog says which problems use a live model.

Eight levels: LLM APIs, Search, RAG, Tool Calling, Agents, MCP, Evals and Production. The curriculum groups them into sixteen tracks; the first four (Model APIs, Context engineering, Retrieval, RAG and grounding) are open now, and the rest are being written.

You can still submit and pass, but the solve earns no Heat, and it is not pushed to your GitHub, so the repo only holds code you wrote.

A coding agent you build on the site: a small Python repo whose harness.py defines run_coding_agent. It starts as the simplest agent that works and is yours to change. Repair problems are solved by it, and only by it; there is no built-in agent.

Test runs it in your browser against a scripted model and a practice repo, free: the check (pass it and repair problems can run on it), a bench of five models that misbehave, and any test files you write yourself. Chat and repair runs use a real model and spend credits.

In your browser, and on your account too when you are signed in, so it follows you between devices. You can also export it as a file and import it elsewhere.

A real program built one tested step at a time, such as a model client or a tool-calling agent. Your passed code from earlier steps runs before each new step, and when every step passes you run the whole program. Some projects build your agent for you as you go.

Yes. Connect GitHub from the panel beside the problem catalog and your solves and passed project steps are committed to a public repo called deep-agnt on your account, under your name. You can push a project or your agent at any time, switch automatic pushes off, or disconnect. Only your own code is pushed, never a problem statement or a reference solution.

Every problem, the Docs, reference solutions, and monthly credits for everything that calls a model. Problems with a scripted model are unlimited.

33× Free's monthly credits, hints, the project tutor and the premium projects, for $18 a month or $144 a year. Both plans run every model call on Gemma 4.

Model calls cost money, so each month comes with credits that they spend. They refill on the 1st. The meter in a problem's toolbar shows how much is left; when they run out, work that needs a model stops where it is and the rest of the site keeps working.

Yes. While your Deep-ML plan is active, Deep-Agnt monthly is $5 off; a yearly plan is $60 off when you have one at purchase.

From Your Premium on the site, at any time. You keep Premium until the end of the period you paid for.

Yes. One account signs you in to both sites. Progress is kept separately, and so are the subscriptions.

Write to info@deep-ml.com from the email on your account to ask for a copy of your data, a correction, or deletion. Because the account is shared, deleting it removes it from Deep-ML as well.

It is coming. Until it opens, you can tell us which companies you want covered from Interview Prep in the menu.