Ahmed Hmeid Contact
All work 04 · AI study buddy for children

Know-It Owl

Knows it all. Makes you work for it. A study owl for Years 5 to 9 that never gives the first answer.

Open know-it-owl.hmeid.uk

What it is

Know-It Owl helps children with maths, science and history homework without doing it for them. The owl asks questions, gives clues, finds a vetted page to read or draws a diagram, and only gives the answer once the child has had a real go. Then it asks them to explain it back and checks about a week later that it stuck. Every message in and out is checked by a separate AI before the child sees a word, and anything worrying reaches the right adult.

  • 422red-team prompts run before every release
  • 1.1%answers leaked on 379 unseen homework questions
  • 1,150+vetted reading pages from 42 sites
  • 200+automated tests

Key features

  • Clues before answers

    The owl asks, hints and points to a page. Parents choose how hard it makes them work, from one clue to about eight tries, and every level ends in an answer.

  • Pictures and proper maths

    Ten kinds of hand-drawn diagram, from fraction bars to circuits, plus fractions and powers set properly and a symbol bar for typing them.

  • Snap your homework

    Take a photo of a worksheet and the owl reads the question. The photo is checked, read and never stored.

  • Explain it back, then a check-up

    The owl asks for the idea in their own words, offers three quick questions, and re-quizzes the topic about a week later.

  • Four owls to choose from

    Olive is warm, Pip is cheeky, Professor Quill is serious and Dot uses simple words. Badges are earned for real effort and dress the owl up.

  • Parents stay in charge

    Child logins without email, subjects and rules per child, alerts when it matters, and a Sunday summary with something to ask at dinner.

In action

The landing page: "Knows it all. Makes you work for it."
A child starts by typing a question, picking a curriculum topic or snapping their homework.
The owl draws fraction bars, then catches a common mistake without giving the answer away.
Once the child has worked it out, the owl asks them to explain it in their own words.
A three-question check-up to finish, with badges awarded in code for real effort.
In history, the owl hands over a vetted page to read, with what to look for, and keeps it in the notebook.
Parents add a child with a nickname only, choose one of four owls, and get a generated login name.
Parents see what their child worked out this week, and something to ask at dinner, without reading every chat.
Rules per child, including how hard the owl makes them work before it gives an answer.
The start screen on a phone.
Diagrams and proper fractions work on a phone too.
On a phone, the notebook slides up with the pages to read and the badges earned.
Under the hood

The parts that aren't obvious

  1. Checked by a second model, not just a prompt

    Telling a model to behave isn't enough when the reader is a child. Every reply Claude drafts is held back and checked by a different model through OpenAI's new Decisions API, which answers typed questions about it, such as whether it suits the child's age or gives the answer away, with a probability for each. Code makes the call from those answers, not the tutor. A reply that fails is rewritten or swapped for a safe hint, and if a check fails or times out, nothing is sent.

  2. The answer is unlocked in code

    Whether the owl may give the answer is worked out each turn from what the child actually did, not left to the model. Saying "I already tried, promise" got the answer in 5 of 15 test chats; after the fix, 1 in 15.

  3. Concerns that can't be lost

    A crisis reaches Childline even over the message cap or when the checks are down. If harm at home is possible, the whole chat is hidden from parents and goes to a reviewer instead.

  4. Measured, with the misses published

    A 422-prompt red team, held-out leak tests and 120 multi-turn chats against an AI playing a crafty child. A Sonnet trial cost over three times as much with no fewer leaks, so cheaper checks on Haiku shipped instead.

  5. Private and budget-capped

    Its own database role on a shared Supabase project, with no Data API and no secret key. Chats are deleted after 30 days, and AI spend stops at £2 a day.

Built with

App
  • Next.js 16
  • React 19
  • TypeScript
  • Tailwind CSS v4
  • KaTeX
  • Rough.js
AI
  • Claude Haiku 5.5
  • Claude Sonnet 5.5
  • OpenAI Decisions API
  • OpenAI moderation
Data
  • Supabase Postgres
  • Supabase Auth
  • Resend
  • PostHog (no cookies)
Quality and hosting
  • Vitest
  • Red-team and leak evals
  • Docker
  • Coolify