KaizenEDU

A tutor, a daily practice plan and a school organizer for families. Starting with US students in kindergarten through 9th grade in math, science and English (including rhetoric), with a parent view that shows truthfully what happened. Final home: kaizenedu.net.

It serves four jobs in one product: help right now (homework, a test tomorrow), daily practice (an at-home tutoring center), homeschool (skill maps, lessons, reading log, dated records) and staying on top (school calendar import, test prep that schedules itself).

Status: works end to end in the browser on demo data. The AI tutor and lesson writer are built and switch on when the deployment has an AI provider; without one, a demo tutor uses vetted hints and worked examples. Server accounts, the database and the children's-privacy consent flow come next. What is done, what is next and the latest test results: docs/STATUS.md (the build queue) and docs/ROADMAP.md. Codex's system audit of 2026-10-07 lists where the pieces don't yet connect; its fixes are in the queue.

Phone: Today (K) · a K practice set · Today · Talk · records

Run it

Requires Node 20.9+.

npm install
npm run dev

Open http://localhost:3000 → Get started → create a family account (stored only in your browser) → add learners → pick what you want help with → open a learner's Today.

To turn on the AI tutor and lesson writer, put one of these in apps/web/.env.local (never commit it) or in the Vercel project's environment:

ANTHROPIC_API_KEY=…            # Anthropic directly (preferred); on Vercel also set KAIZEN_AI=anthropic
AI_GATEWAY_API_KEY=…           # or Vercel AI Gateway, Anthropic models only
KAIZEN_AI=gateway              # or, on Vercel, the deployment's own OIDC token through AI Gateway

See apps/web/.env.example. On Vercel an Anthropic key is ignored unless KAIZEN_AI=anthropic is set, because the project still holds keys from earlier attempts; requests always go to Anthropic's own API.

Settings shows which one is live. npm run verify runs lint, type check, the unit tests and a production build. npm run e2e runs the browser journeys and accessibility checks at desktop and phone sizes (CI=1 npm run e2e when a dev server is already running). Current counts are in docs/STATUS.md.

What's here

Path	What
apps/web/src/practice	The skill map (134 skills: 76 math, 30 English, 28 science, K–9, CCSS/NGSS codes), seeded problem generators with hint ladders and worked steps in EN/ES, a safe algebra parser and the answer checker. Every answer is checked by code.
apps/web/src/learning	The learning engine: skill status (practicing → check ready → passed one check → proved → needs a refresh), level stepping, review spacing, stuck and overdue flags, placement, set building.
apps/web/src/planner	Today's plan, calendar dates, .ics import/export, reading pasted school text, matching school words to skills.
apps/web/src/lib/ai	The tutor (prompts by grade band, tools, safety screen), the lesson writer with quality gates, AI practice, document reading, the family note.
apps/web/src/resources	Named free sources per skill and topic (PhET, Khan Academy, OpenStax, NASA, Project Gutenberg, Purdue OWL…), links checked.
apps/web/src/catalogue	Hand-written K–9 courses for the lesson stage.
modules/	Everything reusable from the earlier attempts — parked, not built.
docs/	Decisions, specs, plans, roadmap, status; docs/history/ for earlier attempts.

The rules it keeps

• Practice is not proof. Answers are on your own, with help, or not yet. A skill is proved only by two short checks with no help available, on fresh problems, on different days at least six days apart, the first at least 48 hours after the last help.
• Code checks answers; AI talks. The tutor asks before telling, never gives a live answer before a try, and calls tools for hints, checking and worked examples. No answer key is in its prompt.
• Safety before any model. A crisis or abuse disclosure gets a fixed referral (988, Childhelp) and a note for the family. Names never reach a model. Transcripts are visible to grown-ups.
• Engagement earns its place. A game-like mechanic is fine when it serves the learner's own goal, has evidence that it helps at that age, comes in its smallest form, and never fakes progress or leads anywhere else. Nothing nags a child. With every option off, it is a calm, clean teacher (1.0 plan §2.10).