The reframe
Learnify 1.0 (2023 hackathon) generated static AI courses — a loop that's now a commodity. Learnify 2.0 stops generating content and starts generating the experience: an agentic tutor (Claude on AWS Bedrock) that runs live adaptive sessions where every screen — explanations, checks, flashcards, diagrams, exercises — is streamed as A2UI declarative JSON and rendered from a safe client-side widget catalog. A persistent mastery engine (concept graph + spaced repetition) makes it a daily habit, Supermemory gives the tutor cross-session memory of how you learn, and attached sources (PDF/URL/YouTube) ground teaching with citations.
Stack: Next.js 15 + TypeScript, @anthropic-ai/bedrock-sdk (strict tool-use emission of A2UI), @a2ui/react, Neon Postgres + Drizzle, Better Auth + demo mode, Vercel.
Phases
Phase 1 — Vertical slice (demo-able core)
Phase 2 — Full catalog + hardening
Phase 3 — Mastery + memory (the moat)
Phase 4 — Learn from anything
Phase 5 — Ship it
Stretch
Final acceptance
A deployed URL where an incognito visitor clicks “Try the demo”, enters any topic, and is in a live session where the tutor streams interactive widgets, adapts to wrong answers, remembers how they learn across sessions, builds a mastery graph on /today — and with a YouTube link attached, teaches with citation chips that jump to the exact quote.
Key risks being managed
A2UI v0.9 churn (renderer isolated to one file, hand-rolled fallback planned) · malformed LLM JSON (strict tools + atomic block rendering) · demo cost (budgets, caching discipline, model downgrade) · Bedrock gaps (no Files API → base64 PDFs ≤ 20MB; client-side refusal handling) · YouTube transcript fragility (fallback chain + manual paste).
The reframe
Learnify 1.0 (2023 hackathon) generated static AI courses — a loop that's now a commodity. Learnify 2.0 stops generating content and starts generating the experience: an agentic tutor (Claude on AWS Bedrock) that runs live adaptive sessions where every screen — explanations, checks, flashcards, diagrams, exercises — is streamed as A2UI declarative JSON and rendered from a safe client-side widget catalog. A persistent mastery engine (concept graph + spaced repetition) makes it a daily habit, Supermemory gives the tutor cross-session memory of how you learn, and attached sources (PDF/URL/YouTube) ground teaching with citations.
Stack: Next.js 15 + TypeScript,
@anthropic-ai/bedrock-sdk(strict tool-use emission of A2UI),@a2ui/react, Neon Postgres + Drizzle, Better Auth + demo mode, Vercel.Phases
Phase 1 — Vertical slice (demo-able core)
Phase 2 — Full catalog + hardening
Phase 3 — Mastery + memory (the moat)
Phase 4 — Learn from anything
Phase 5 — Ship it
Stretch
Final acceptance
A deployed URL where an incognito visitor clicks “Try the demo”, enters any topic, and is in a live session where the tutor streams interactive widgets, adapts to wrong answers, remembers how they learn across sessions, builds a mastery graph on
/today— and with a YouTube link attached, teaches with citation chips that jump to the exact quote.Key risks being managed
A2UI v0.9 churn (renderer isolated to one file, hand-rolled fallback planned) · malformed LLM JSON (strict tools + atomic block rendering) · demo cost (budgets, caching discipline, model downgrade) · Bedrock gaps (no Files API → base64 PDFs ≤ 20MB; client-side refusal handling) · YouTube transcript fragility (fallback chain + manual paste).