Adaptive AI Scaffold

A research platform for designing AI writing tools that preserve student reasoning, encourage engagement with course materials, and make learning processes visible to teachers.

Live demo: https://kishormorol.github.io/adaptive-ai-scaffold/

---

Overview

Generative AI tools are increasingly used in academic writing. While they improve productivity and fluency, they often function as answer engines, reducing reasoning effort and obscuring how students arrive at their final work.

This project explores an alternative paradigm:

AI as an adaptive scaffold that supports thinking rather than replacing it.

Instead of providing direct answers by default, the system controls when and how assistance is given based on student input and effort.

---

Screens

✍️ Write

The main student workspace. Includes an assignment header, reasoning gate, and paper-styled editor. The "Ask AI" button opens an inline guided flow — not a chat.

🤖 Ask AI (inline flow)

A 3-step guided interaction:

1. Context — student describes what they've tried (enforced by gate strictness)
2. Kind of help — tiered options ordered by how much thinking they leave to the student:
   • Hint → Critique → Explain → Outline → Full answer (last resort, requires reflection)
3. Response — grounded in course materials, with source cards and echo-back of the student's own draft

🔍 Trace

Before/after diff view of the student's draft across three snapshots. AI-authored spans are highlighted and hoverable — each links back to the AI event that produced it. Includes an interaction timeline and source grounding footnotes.

👩‍🏫 Instructor

Class roster with initial reasoning histogram, flags (students who used last-resort help), and AI usage breakdown. Clicking a student opens their trace with auto-surfaced "Key Moments" interpretation.

---

Reasoning Gate

Controls how strictly students must write before AI unlocks. Configurable via the top bar or the § Tweaks panel:

Mode	Behavior
Soft	Suggested minimum; skippable
Medium	Word threshold required; reason to skip
Hard	AI disabled until threshold met; requires structured reflection before any AI use

---

Tech Stack

• Frontend: Next.js + CSS variables (paper/lab-notebook aesthetic)
• Deployment: GitHub Pages (static export)
• Backend (planned): FastAPI
• Database (planned): Supabase / PostgreSQL
• AI (planned): LLM API + Retrieval-Augmented Generation (RAG)

---

Local Development

npm install
npm run dev

Open http://localhost:3000.

---

Project Structure

components/
  data.js          # Assignment, sources, stages, class roster
  primitives.jsx   # Icons, TopBar, FigCaption, SourceSnippet, MiniBar
  writing.jsx      # Student workspace + reasoning gate
  askai.jsx        # 3-step Ask AI flow + HELP_KINDS
  trace.jsx        # Diff view, timeline, actions panel
  teacher.jsx      # Instructor dashboard + student detail
  tweaks.jsx       # Gate strictness panel
pages/
  index.jsx        # App shell + screen routing
  _app.jsx         # Global CSS import
styles/
  globals.css      # Design system (paper tones, typography, animations)
project/           # Original HTML/CSS prototype (reference)

---

Research Motivation

Current AI tools optimize for output quality and efficiency but neglect:

• reasoning effort
• engagement with learning materials
• transparency in the learning process

This project investigates:

Can AI tools be designed to support writing without replacing learning?

---

Planned Study

Three conditions will be compared:

1. Standard chatbot
2. Static scaffolded AI (limited assistance)
3. Adaptive scaffolded AI (this system)

Measures

• Writing quality
• Reasoning depth
• Source engagement
• AI usage behavior
• Teacher understanding of student learning