Montenova
Montenova gives parents a say in what their children watch. An AI picks YouTube videos around each child's interests and puts activities and breaks between them, so screen time follows a routine instead of running until someone stops it.
- Context
- Selected for the European Innovation Academy, a three-week intensive accelerator for international teams building a startup from scratch, with mentorship from industry professionals.
- Role
- Team CTO, leading product development.
The problem
Screen time for kids is either unstructured (a tablet with no limits) or rigid, a fixed timer with no room for what the child is actually doing. Parents wanted something in between: real oversight, but a system that adapts to the actual content, not just a countdown.
What I built
A SaaS platform where parents build a screen time program: blocks of curated video pulled from the YouTube Data API, mixed with guided breathing and movement breaks, with the schedule adapting elastically to how long each video actually runs. Built on Next.js, TypeScript, Supabase (Postgres and Auth), Tailwind, and the Claude API for the AI-assisted scheduling layer. Deployed on Vercel with continuous delivery from GitHub, a custom domain and SSL.
- AI-driven programming: the Claude API selects and sequences videos against a child's preferences, and generates activities and breaks between them.
- Multi-profile households, with a who-is-watching flow, per-child preferences, onboarding and access codes.
- Installable as a progressive web app, with a service worker and install prompt.
- Internationalised from the start with next-intl, and deployed on Vercel.
Decisions & trade-offs
Directing the build, not typing it
As CTO I made the calls rather than writing every line by hand, directing development with Claude Code and approving or rejecting every technical proposal before it shipped.
A PWA over a native app
A PWA instead of a native app, to skip app store review cycles.
Feature flags over deleting code
A feature-flag system, so we could switch off half-finished work without deleting code.
What I learned
That reviewing AI-generated code needs to be proportional to risk, not uniform.
Every feature had to pass end-to-end tests against a real production build and a real database, not just a dev environment, and that discipline caught things a lighter review would have missed: a database security policy that had only been half applied, a colour-contrast accessibility issue, and a race condition in UI state. The tests weren't a checkbox, they were how I found out where the AI-written code and my assumptions had actually diverged.
Tech stack
- Next.js
- TypeScript
- React
- Supabase
- Claude API
- Tailwind CSS
- next-intl
- Playwright
- Vercel