Sidekick

What it is

Most AI courses teach children to use AI. Sidekick teaches them to build it — and to be properly sceptical of it. From the first hour the model is the specimen on the bench, never the teacher: every lesson opens with the instructor breaking a model in front of the room, and the concept of the week arrives as the repair. Children see the machinery the industry hides — the context window filling, the retrieval that found the wrong paragraph, what a token costs — because the instruments are on screen at all times.

Each child builds one continuous artefact across the whole programme. In Level 1 it learns their voice, reads their documents and passes tests they wrote themselves. In Level 2 it reaches real tools over real protocols, researches the web under guardrails, and joins a small society of classmates' agents. In Level 3 they engineer it properly — measured, optimised, audited for bias — and finish with a capstone of their own choosing, presented publicly. Nothing is a throwaway exercise; three years of work is one growing thing they own.

What a child builds

• Level 1 — AI Builder

  How do I make a model do what I want, reliably?

  The year the Sidekick comes to exist. Prompting as engineering rather than typing; a job description that survives attack; the context window as a budget; examples, reasoning, structure and honest testing. Then the Sidekick learns things only this child knows — their documents, chunked, embedded and drawn as a graph they can read — and gains hands: tools written in the child's own first Python. The blocks retire in week 25 and the year ends with a working, tested, published-to-the-family agent and the AI Builder title.

• Level 2 — AI Engineer

  How does my agent reach the world, and other agents?

  The year the Sidekick stops being a program the child runs and becomes a participant in a wider system. It reaches real tools over MCP — including servers the children write for each other — researches the open web under guardrails, learns to act on a schedule with a kill switch, and joins a society of classmates' agents: delegation, trust, a class agora, and the attacks that societies of agents make possible, met in a bounded arena. The year closes with the software-engineering term — tests, git, review, CI, deployment, observability — and the AI Engineer title.

• Level 3 — AI Architect

  How do I build something real, and prove it works?

  Depth, judgement and a capstone. Retrieval done properly — hybrid search, graph traversal, multi-hop, freshness — building a real knowledge product for a real user with measured quality. Then the judgement term: the model landscape, local models, the prompt/retrieve/fine-tune decision, latency, cost at scale, reliability, bias, and a structured debate with no prescribed conclusion. The final term is the capstone: a real problem, a design document, six build sprints, real users, handover — and a public Demo Day, the AI Architect title, and the programme certificate.

What you see, as a parent

You see everything. Every message between your child and any model is readable in your portal, updates arrive after every session, and a report you can hold lands every half term. There are no marks, no leaderboards and no streaks — progress is shown as your child against their own past, in their own words. You buy six weeks at a time; leaving after any block leaves a complete thing, and returning picks up exactly where they stopped.

What we are promising

A child who leaves genuinely proud of something they made — and who can explain exactly how it works, which is the part no demo can fake.

Enrol