Web

An adaptive-learning runtime that builds a per-learner concept graph from any source text, then drives question selection against it, inserting newly-detected prerequisite gaps as live graph nodes mid-session.

video_2x_small.mp4

---

What it does

A learner gives Web a goal and a source (paper, syllabus, chapter). The system runs a brief Socratic intake, then synthesizes a prerequisite DAG of concepts grounded in the source, lays it out as an interactive graph, and runs adaptive practice over it. Each attempt updates the learner's proficiency, schedules the next review via SM-2, and — when the LLM classifies a wrong answer as a prerequisite gap rather than a misconception — proposes a new concept node that, once accepted, is inserted into the graph with inferred edges and pre-generated questions.

The loop closes inside a single session: a wrong answer can grow the graph it's drawn from. Mastered concepts trigger an extension proposal so the graph also expands upward toward the learner's stated goal.

---

Technical notes

• Concept graph is first-class and multi-tenant. UnitGraph rows share Concept rows through GraphMembership rather than duplicating them; mastery state lives on the concept, not the membership. Cross-graph extensions are tracked by GraphLink. (adaptive-tutor/prisma/schema.prisma)
• Mastery is a two-clock model. SM-2 decides when to review; an Elo-style update decides how much a single attempt moves the needle. Backward prerequisite-confidence propagation lets a correct answer on a high-level concept lift the priors on its dependencies. (adaptive-tutor/src/lib/algorithms/sm2.ts, adaptive-tutor/src/lib/algorithms/questionSelector.ts)
• Question selection scores three signals. priority = uncertainty × graph-centrality × prerequisite-readiness. Difficulty within a concept targets proficiency + 0.15 — Vygotsky's zone of proximal development as a one-line rule.
• LLM is in the loop, not on top of it. Curriculum-graph synthesis emits structured JSON validated by Zod and cycle-checked via Kahn's algorithm before persistence; an error classifier grades free-response answers into CORRECT / MINOR / MISCONCEPTION / PREREQUISITE_GAP and writes to a gap-detection ledger that drives live graph mutation.
• Provider-agnostic LLM transport. Vercel AI SDK v6 with Anthropic-compatible endpoint pluggability — defaults to MiniMax, override ANTHROPIC_BASE_URL for native Anthropic, switch packages for OpenAI.

---

Stack

Next.js 16 (App Router, Turbopack) · React 19 · TypeScript · Tailwind v4 · Prisma 7 + libSQL / SQLite · Vercel AI SDK v6 · Zustand · @xyflow/react · d3-force · Zod · Vitest.

---

Run it

cd adaptive-tutor
cp .env.example .env       # fill in ANTHROPIC_API_KEY (or MINIMAX_API_KEY)
npm install
npx prisma db push          # initialize SQLite at prisma/dev.db
npm run dev
open http://localhost:3000

---

Architecture

See ARCHITECTURE.md for the closed-loop diagram — curriculum-graph synthesis → question selector → attempt + LLM evaluator → proficiency / SM-2 / gap detection → graph mutation → back to the selector.

---

Layout

adaptive-tutor/        — Next.js app (all runtime code)
  src/app/(tabs)/      — chat, learn, graph
  src/app/api/         — 30+ route handlers
  src/lib/algorithms/  — sm2, questionSelector, graphValidator, forceLayout
  src/lib/prompts/     — phase-specific LLM prompts
  prisma/schema.prisma — concept-graph + evidence ledger
  scripts/             — seed and debug scripts
  docs/internal/       — design notes
ARCHITECTURE.md
LICENSE

---

Where it is

v1 was the project that won Yale's CS hackathon. Current work is a learning-outcomes study in collaboration with Yale faculty; a handful of edtech teams are looking at it in parallel.