← All work
LiveZazi iZandi · Masinyusane

An offline-first literacy app with an AI coach in every pocket

The app young education assistants use to teach children to read in South African schools, now with an AI coach that knows each child's progress and says what to teach next.

Status: AI Coach rolling out to its first education assistants

teaching sessions in 30 days
9,400+
teaching sessions in 30 days
education assistants active in that month
190+
education assistants active in that month
child attendances recorded
52,000+
child attendances recorded
automated tests before every release
8,000+
automated tests before every release
Zazi iZandi app: the AI Coach suggesting which letters to teach next and which children need extra turns
Zazi iZandi app: a live teaching session with today's suggested letters, a timer and attendance
Staff dashboard: a performance map plotting each education assistant by teaching quality and session dosage

The problem

Zazi iZandi trains young people as education assistants who teach small groups of children to read, in schools where connectivity is patchy and phones are often low-end. They needed to know exactly what to teach each group next, and managers needed to see where coaching was needed, without waiting for paper or a third-party tool that couldn't keep up.

What I built

  • A React Native app that works fully offline: GPS clock-in, session recording, letter assessments, and automatic grouping of children by level, syncing in the background when signal returns.
  • Staff dashboards on the web: sessions, attendance, programme fidelity, and a nightly coaching queue that shows managers which assistants need support.
  • A release process built for a fleet of phones in the field: thousands of automated tests on every release, and database changes that stay compatible with older app versions still installed.

The AI inside

  • The AI Coach: a chat tab that knows the assistant's own classes, groups, letter tracker and assessments, and answers with clear cards: which letters to teach next, which children need extra turns.
  • Privacy and safety by design: children's names are replaced with codes before anything reaches the model, every answer is labelled as AI-written and can be reported, sensitive disclosures get a fixed safe reply, and daily limits keep costs predictable.
  • The Coach's teaching advice is checked against the app's own teaching engine by an automated test harness, so the AI and the app never disagree.

Where it stands

The app replaced the previous field tool and now carries the programme: 9,403 teaching sessions by 194 education assistants in a single 30-day window (September 2026). The AI Coach has shipped in production and is being switched on for its first trusted education assistants.

Expo / React NativeSQLite (offline-first)Supabase Postgres + RLSDjangoNext.jsOpenRouter LLMs

Want something like this?

Tell me what you're trying to do. If I can build it well, I'll show you how; if not, I'll point you somewhere better.