The AI Engineering track

The track

You are preparing for Track 3 of the Entry-Level AI Skills framework, AI Engineering: you build a product whose main feature is a model somebody else trained. Switch the model off and the product stops working. Six weeks, eleven areas, one repository that grows week by week into something you can deploy, measure and defend in an interview.

#	Cluster	Area	Week
1	Foundations	Software foundations for AI work	0 (assessed, not taught)
2	Foundations	Working with a model as a component	1
3	Foundations	Context engineering	2
4	Retrieval and agents	Retrieval, done properly	2
5	Retrieval and agents	Grounding and honest answers	3
6	Retrieval and agents	Agents and tools	4
7	Retrieval and agents	Connecting to real systems (MCP)	4
8	Trust and shipping	Evaluation	3
9	Trust and shipping	Observability and cost	5
10	Trust and shipping	Security and guardrails	5
11	Trust and shipping	Shipping and proving it	6

The two areas that decide most hiring outcomes are Retrieval (Week 2) and Evaluation (Week 3). They get the most hours.

How a week works

Every week has the same shape: Read, See, Build, Measure, Decide, Submit. You never scaffold a new project after Week 0. Each week names one lesson; every build item is core (needed for it, checked by the gate) or stretch (shown to your mentor, never blocking).

flowchart LR
  A["Read
the concept, 1 h"] --> B["See
the scripts, 0.5–2 h"]
  B --> C["Build
5–7 h, any assistant, logged"]
  C --> D["Check
make check WEEK=N"]
  D --> M["Measure
make evidence WEEK=N"]
  M --> E["Decide
decisions/week-N.md"]
  E --> F["Submit
PR: checks"]
  F --> G["merge
hidden evaluation, mentor's brief"]

Read: one page, one sentence you must be able to say back in your own words, a few diagrams and a short reading list. Write the sentence in reflections/week-N.md, Q1.
Run: one or two small scripts and a few files to read. Change nothing yet.
Build: what the week asks for, specified as behaviour that must hold, not as a list of functions. Use any assistant you like, and log it in ai-use/week-N.yml as you go.
Check: make check WEEK=N, as often as you like. Green is the goal, not the end.
Measure: make evidence WEEK=N runs the week's measurements on a real model and keeps them in evidence/week-N.json, with the commit, provider and model.
Decide: decisions/week-N.md, the one decision of the week, with the numbers from the evidence file. The gate checks that they match.
Submit: a pull request from week-N to main, 24 hours before a session. The checks run. Merging it is what marks the week done: within a day the programme collects the data for your mentor's brief from what you merged and, from week 2, runs its hidden evaluation on questions you never see.

Your route (start, core or pro) is set by your mentor at discovery from two questions, in order: can you do the basic work of the track at all; and do you check what you produce. It holds for the programme, with one correction point at session 2. The route changes what you are given, never what is measured:

	start	core	pro
Exercises	faults named, a worked example given (`weeks/N/routes/worked_example.py`)	faults present, not located, count not given: each exercise file is complete with mistakes planted	no helpers, one real constraint added (`weeks/N/routes/pro.md`)
Pace	slow, mentor-supported	compressed	self-directed
Finished piece	modest and completed	full	harder, made public

You make your repository from your route's template, so it holds that route and no other; each week's routes/ page says what it means for the week. If the route is corrected at session 2, the programme opens a pull request on your repository that swaps the files for the weeks still ahead.

Sessions and gates

Mentor time is four sessions, and nothing else except up to two fifteen-minute unblock calls. That is the framework's rule, and it is what makes the route matter: the sessions are where the hours go, so what they are spent on depends on where you started.

Session	When	Length	What happens
1. Discovery	Week 0	45 min	An observed task sets your route and takes the starting measure. The finished piece is agreed
2. Direction	end of Week 1	45 min	Route confirmed or a correction recommended. Unseen items with your assistant; your decisions without it
3. Observation	end of Week 3	60 min	You explain the gap between your score and the hidden score, diagnose a hidden failure with your assistant, then defend your threshold and judge without it
4. Defence	end of Week 6	60 min	Demo; three decisions defended with tools closed; an unseen change with your assistant; the starting measure again

From session 2, every session has a half where you drive your own AI assistant on something unseen while your mentor watches how you steer it, and a half with tools closed. Interviews now have both kinds of round.

Weeks 2, 4 and 5 have no session: the gate, the self-tests and this page are your feedback, and your mentor passes on the hidden evaluation's totals. Your mentor reads a one-page brief built from each week you merge (they do not read your diff) and gives a one-click verdict: on track, something to probe at the next session, or an offer of an unblock call.

Unblock calls: up to two, fifteen minutes each, on request, after you have attempted the thing yourself. Send what you tried with the request.

Where the hours go, by route:

Route	Sessions 2 and 3 concentrate on	Pace
start	the basics: Weeks 1–2, getting the model layer and retrieval working at all	slow, mentor-supported; use both unblock calls early
core	quality: Weeks 3–5, checking what you produce: evaluation, guardrails, traces	compressed
pro	evidence: Week 6, the deployed piece, the write-up, the defence	self-directed

Three layers of evidence, from cheapest to most human. None of them is self-reported.

Layer	Who	What it measures
The gate: `make check WEEK=N`	CI, on every pull request, with a fake model so no key is needed	Lint, strict types, that week's tests under two fake providers, the records, and from Week 3 the evaluation gate, from Week 5 the attack gate
The hidden evaluation	The programme, on each week you merge from week 2, beside its reference solution: every item on the fake model (with floors), a sample on your `model.json` model (no floor)	About 150 questions and 20 agent tasks written for your product, never shown to you: your score is a share of the reference's, and your mentor brings you the totals
Your records	You write; the gate checks them; your mentor's brief shows them	The decision and the numbers behind it, what an assistant did and what you checked, your reflection
Your mentor's judgement	Sessions, against one rubric	How you steer an assistant on something unseen, and whether you can explain your own work without one

Work reaches your mentor a day early, so the session is not spent finding out what is there. You talk more than your mentor, because otherwise the hour drifts into them solving the problem.

Understanding sets a ceiling: at the Defence your mentor judges whether you can explain your own work, and that judgement caps what the programme records; it never adds to it. You are always shown what you produced and what it missed. You are never shown the measure that goes to the programme, and neither is this page. The self-tests here are yours alone.