Fast enough to be useful in class
An explanation that takes ten seconds arrives too late. On average it is there after about three seconds.
Not simply trusting the AI
Every answer is checked before it appears on the page: its format, its evidence and its suitability for young users. For questions about a whole subject, the model is given matching passages from handwriting and PDFs as evidence.
Keeping cost and abuse under control
The AI requests go through my own server, written in TypeScript on Cloudflare Workers. App Attest ties every request to a real device, the quotas apply across all devices of a subscription, and cost is capped in advance. Nobody needs a separate account for this.
Smooth writing on full pages
Otriva saves once per writing pause instead of after every stroke. Sync between iPad and iPhone runs through the user's private iCloud.