AI Engineering Programme Start here Weeks Playground Open Codespace

Discover → Build · AI Engineering

One service, six weeks, numbers you can defend.

A support desk works the same kinds of ticket every day, from a manual that exists in two versions worded almost identically. You build the agent that answers from the right one with a citation or declines, looks up the customer's own data, and acts only when a person approves. New here? Track 3 in ten minutes.

route start1 of 7 weeks completeupdated 18 Oct 15:20 UTC · fe45508

Howyou get there

01

Getting ready

Open the workspace, read the client's brief, and prove the foundations.

Week 0 complete · Open

02

The six weeks

One agent, changed every week: build, measure on a real model, decide, merge.

0 of 6 milestones complete · Open

03

Your evidence

Measurements, a published build, and one page on what it did for the business.

Ready at week 6 · Open

Weekswhere you are

0Foundations gate
all 8 checks pass
checks pass›
1The model as a component, and its first tools
1 of 16 checks pass
submitted›
2Context and retrieval, as the agent's main tool
17 of 34 checks pass
checks fail›
3Grounding and evaluation, of answers and of paths
5 of 18 checks pass · citations 0% valid (fake model)
checks fail›
4Agents that act, and where agency pays
7 of 30 checks pass
checks fail›
5Observability, security and privacy
6 of 30 checks pass · 8 of 8 attacks blocked
checks fail›
6Shipping it, presenting it and proving it
10 of 14 checks pass
checks fail›

Referencewhen you need it

Built 18 Oct 15:20 UTC from main at fe45508. It rebuilds on every merge to main and every push to an open pull request, in about ten minutes, and your browser can show the previous build for up to ten more: reload then. A push to a branch with no pull request does not rebuild it. While you are working on a branch your terminal is ahead of it: make next and make check are the live answer, and this is the record. Nothing here is typed by hand. Numbers measured against the test model and the lexical embedder prove the pipeline holds, not that answers are good — each week says which it ran. progress.csv · progress.json