Curriculum-grounded AI teacher for school students

Voice-First AI Tutor

A personal AI teacher that speaks and listens, teaches from the student’s own textbooks, draws on a whiteboard and tracks what they have mastered.

AT A GLANCEEDUCATION
Apps
Student PWA, parent dashboard, admin console
Teaching
Live bilingual voice, or streamed text
Grounding
Hybrid vector and full-text retrieval with citations
Built with
Next.js, Fastify, PostgreSQL + pgvector
11 MODULESBUILT AND SHIPPED

The problem

What it replaces

General-purpose chatbots hand children answers, invent facts that are not in their syllabus and give parents no idea what was learned. Schools and tutoring businesses want AI help that stays inside the curriculum, teaches rather than tells, and is safe for children.

This tutor ingests and publishes curriculum through a reviewed pipeline, retrieves from it with citations, and teaches in live voice lessons or streamed text — Socratically, with schema-checked visuals, mastery tracking and a parent dashboard. The model only requests actions; the server decides what happens.

Who uses it

Students
One-tap voice or text lessons on a phone or tablet, handed over by a parent with an optional PIN.
Parents
Register with consent, create child profiles and follow progress on a dashboard.
Curriculum admins
Upload or fetch textbooks, review extracted structure, and publish.

Features

What the system does

01

Child-safe accounts

Parent registration with consent, child profiles, scoped learner tokens and a parent PIN — no child email or phone number is collected.

02

Curriculum as data

Board, class, subject and chapter are records, not code, with admin authoring and an audit log.

03

Reviewed ingestion

Upload or fetch from an allow-list, extract page by page, detect structure, then edit, approve and publish behind a gate.

04

Grounded retrieval

Structure-aware chunks, vector and full-text search fused and reranked by curriculum proximity, with citations and a grounded flag on every answer.

05

Live voice lessons

Real-time voice with barge-in and reconnection, falling back to speech-to-text, model and text-to-speech when needed.

06

Teaches, doesn’t tell

Layered prompts drive a Socratic teaching style; the model asks for actions and the server carries them out.

07

Safe visuals

Thirteen schema-validated visual types, such as step-by-step equations, rendered by the app — the model never supplies HTML or SVG.

08

Animated teacher

An animated mascot with expressive states and lip-sync to the voice.

09

Mastery tracking

Knowledge tracing with retention decay and misconception clustering, plus hints, generated questions and practice sets.

10

Parent dashboard

Progress written constructively, a rule the test suite enforces.

11

Security by design

A route-by-role authorisation matrix, prompt-injection tests and a data-retention job.

Architecture

How the pieces fit together

A monorepo that keeps the real-time voice gateway separate from the web app, and keeps the AI model on the server side of a strict tool boundary.

Apps

  • Student PWANext.js · audio worklets
  • Parent dashboardNext.js
  • Curriculum consoleNext.js

Platform

  • Tutor APIFastify
  • Voice gatewayWebSocket
  • Teaching enginetool boundary
  • Retrievalhybrid + rerank
  • Ingestion workerjob queue

Data

  • PostgreSQL 16pgvector + full-text
  • Job queuein Postgres

Connected services

  • AI modelsteaching, live voice, speech, embeddings
The browser never talks to the AI provider. Every model call goes through the server, which checks the request, retrieves from the published curriculum, and executes only the actions the model is allowed to ask for.
  1. PublishTextbooks are ingested, reviewed and published as structured curriculum.
  2. AskA student speaks or types; the server retrieves the relevant pages.
  3. TeachThe tutor guides them with questions and visuals, citing the book.
  4. TrackMastery updates per concept and the parent sees progress.
Built with
  • TypeScript
  • Next.js
  • Fastify
  • PostgreSQL
  • pgvector
  • Drizzle
  • Zod
  • Vitest

Built before, fitted to you

Start from an architecture that has already shipped

We have delivered this kind of system, so yours starts from a design that is known to work, not a blank page. Your workflow, your terminology and your approval chain shape what we build on top of it.

Enquire Now