Fawna
Snap a photo of your child's maths homework. Fawna marks it and tells you which skill to practise.
Live: getfawna.com
The problem
A parent sees "6/10" and doesn't know what to do next. More of the same questions won't help if the child misunderstood one step. Fawna finds the step that broke and names it as a skill, so practice can target it.
What works today
- A parent uploads a photo of a worksheet.
- The photo is shrunk and cleaned (location data removed) in the browser, then again on the server.
- An OpenAI model reads the handwriting and marks every question: correct, incorrect, blank or unreadable.
- Every wrong or blank answer gets a skill from a fixed list of 116, based on the National Curriculum for England (Years 1–6).
- The page shows the score and each question, marked like a real worksheet.
On the live site, uploads need a demo code. Anyone can click Try a sample worksheet to see a saved result.
How the AI part works
- Fixed skill list, not free text. The AI must pick from our list, so the same mistake always gets the same name and can be tracked over time.
- The reply is checked by a schema. OpenAI Structured Outputs with a Zod schema. A wrong answer without a skill is rejected. So is a made-up skill.
- The AI solves before it judges. It writes the correct answer before deciding if the child's answer is right.
- Code counts, the AI doesn't. Totals are added up in code.
- Tested against a golden set. Worksheets I marked by hand, with the right skill for each mistake. Every change to the model or instructions is checked against it.
The best result so far: sending the AI what each skill means (not just its id) took skill picks from 0/2 to 2/2, for about 2,000 extra tokens per call, most of them cached. See ADR-017.
Privacy
These are photos of children's work.
- Photos are never stored. They're sent to the AI and dropped.
- Location data is removed before the photo leaves the phone, and again on the server.
- The child's name is never sent to the AI.
- Logs hold counts only, never a child's answers.
The plan
- Better marking: grow the golden set to 15–20 real mistakes.
- Accounts: parents log in, and the child's school year comes from their profile.
- Understand, then practise: for each skill gap, show the child's answer next to the right one, ask one check question, then give practice. Facts get recall practice, methods get a worked example, ideas get a picture first.
- Spaced review: bring skills back a few days later.
- English: vocabulary worksheets and flashcards.
Full list in docs/progress.md.
The look
Fawna should feel like a warm kitchen table at homework time, not a dashboard.
- Every colour is taken from the hero painting: cream, linen, espresso, lamp amber, terracotta.
- The upload page looks like paper on a desk, and results look like a marked worksheet: ticks, crosses and the child's answer in handwriting.
- Motion is kept small. Only three things move, and each one has a reason.
See ADR-018 and ADR-023.
Engineering decisions
Every big choice is written up in docs/decisions.md: the options, why I picked one, the risks, and how I tested it. Good places to start:
| Decision | Why it matters |
|---|
| ADR-006 Skill names come from a fixed list | Makes results trackable over time |
| ADR-012 Four result shapes | The schema rules out impossible answers |
| ADR-013 Fact, method or idea | Decides what kind of help a child gets |
| ADR-014 Layered error handling | Parents see "try another photo" vs "try again later" |
| ADR-017 Send the skill list | 0/2 → 2/2 skill picks, with token costs measured |
| ADR-022 Demo code before login | Going public without risking the AI bill |
Also: how I tested image shrinking and the golden set.
Tech
Next.js 16 · React 19 · TypeScript · Tailwind CSS 4 · OpenAI Responses API · Zod · sharp · Three.js
Run it locally
npm install
npm run dev
Put your secrets in .env.local:
OPENAI_API_KEY=...
DEMO_ACCESS_CODE=...
Then open localhost:3000.
