Skip to content

victorbash400/parchment

Repository files navigation

Parchment

An AI workspace for modern studying.

Next.js Strands Agents Amazon Bedrock Amazon Nova TypeScript

Parchment is a multimodal study application built for the Amazon Nova AI Hackathon. It brings together source-aware chat, math solving, diagram understanding, flashcard generation, and real-time voice tutoring in one workspace designed for students.

Built For

  • Amazon Nova AI Hackathon
  • Focus areas: Agentic AI, Multimodal Understanding, Voice AI

What Parchment Does

Parchment is built around a simple idea: studying should feel like working in an intelligent workspace, not jumping between disconnected tools.

With Parchment, a student can:

  • chat with a general study assistant
  • upload PDFs, notes, images, and text sources
  • ask questions grounded in selected study materials
  • generate flashcards from a topic or source
  • solve math problems with step-by-step explanations
  • use a whiteboard for diagrams and math work
  • ask questions about uploaded diagrams or board sketches
  • switch into a live voice tutoring experience

Feature Cards

Capability Student Experience Core Tech
Source-aware study chat Ask questions over selected notes, PDFs, and images Strands agents, Amazon Bedrock, Amazon Nova
Math workspace Solve problems step by step with a board and computation tool math_agent, mathjs, nerdamer, canvas snapshots
Diagram mode Analyze whiteboards, flows, and biology/architecture diagrams diagram_agent, multimodal input, canvas inspection
Flashcards Generate and review study decks inside the app flashcard_agent, validated flashcard tool
Voice tutoring Speak with the assistant and get live study aids Amazon Nova 2 Sonic, Socket.IO, real-time audio streaming
Multimodal context Mix text, PDFs, and images in one session PDF extraction, selected-source tools, image inputs

Architecture

Parchment uses a Strands multi-agent design inside a Next.js application.

  • The main parchment_agent handles general study assistance and routes specialized requests.
  • Specialist subagents take over for focused tasks instead of forcing one prompt to do everything.
  • The frontend streams model output and tool events from the backend in real time.
  • The voice experience runs through a dedicated server for low-latency speech interaction.
flowchart LR
    U["Student"] --> UI["Next.js Workspace UI"]

    UI --> CHAT["/api/chat"]
    UI --> VOICE["Voice Mode UI"]

    CHAT --> PA["parchment_agent"]
    PA --> MA["math_agent"]
    PA --> DA["diagram_agent"]
    PA --> FA["flashcard_agent"]

    PA --> ST["Selected source tools"]
    MA --> MT["Sandboxed math tool"]
    DA --> CT["Canvas inspection tool"]
    FA --> FT["Flashcard generation tool"]

    CHAT --> BR["Amazon Bedrock"]
    BR --> N2L["Amazon Nova 2 Lite"]

    VOICE --> VS["Voice server"]
    VS --> NS["Amazon Nova 2 Sonic"]
Loading

Agent Structure

parchment_agent

The main study assistant and router.

Responsibilities:

  • handles general educational chat
  • reads selected study sources when needed
  • routes flashcard requests to the flashcard subagent
  • routes diagram and whiteboard requests to the diagram subagent
  • routes quantitative and symbolic problem solving to the math subagent

math_agent

A specialist agent for solving math problems.

Responsibilities:

  • solves equations and quantitative problems
  • uses a sandboxed JavaScript tool instead of relying only on raw model arithmetic
  • explains solutions step by step for student readability
  • accepts canvas snapshots and uploaded math images as visual context

diagram_agent

A specialist agent for visual interpretation and diagram feedback.

Responsibilities:

  • inspects the current canvas snapshot
  • answers questions about diagrams and sketches
  • gives feedback on structure, flow, and interpretation

flashcard_agent

A specialist agent for study deck generation.

Responsibilities:

  • collects topic and card-count requirements
  • generates structured flashcard payloads
  • validates flashcard quality before the UI renders the deck

Amazon Nova and AWS Usage

Parchment is built on AWS with Amazon Nova models.

Core Model Stack

Layer Service / Model Purpose
Main study reasoning Amazon Bedrock + us.amazon.nova-2-lite-v1:0 fast, cost-effective reasoning for chat and subagents
Voice interaction Amazon Nova 2 Sonic real-time speech-to-speech tutoring
Agent orchestration Strands Agents SDK routing, tools, streaming, subagent control

AWS Features Used

  • Amazon Bedrock runtime for model inference
  • Amazon Nova 2 Lite for general and specialist study reasoning
  • Amazon Nova 2 Sonic for conversational voice
  • environment-based AWS region/profile setup for local development

Bedrock / Nova Capabilities Used

  • tool use
  • multimodal understanding
  • real-time voice interaction
  • agentic routing through specialist subagents

Multimodal Workflow

Parchment is designed around mixed study inputs, not just plain chat.

Supported study inputs:

  • PDF documents
  • text-based notes and documents
  • images
  • whiteboard sketches
  • spoken questions in voice mode

How that flows through the system:

  1. A student uploads sources into the workspace.
  2. They activate the sources relevant to the current question.
  3. The frontend extracts text where possible and passes selected context to the backend.
  4. Agents use source tools or multimodal inputs depending on the source type.
  5. The answer is streamed back into the workspace, along with any tool output such as flashcards.

Student-Centered Design

Parchment is built as an education app first.

The product direction is intentionally student-facing:

  • one workspace instead of many disconnected tools
  • support for studying from actual course materials
  • visual and voice-based interaction modes
  • step-by-step reasoning for learning, not just answers
  • flashcards and quiz-style workflows for active recall

Tech Stack

Layer Technology
Frontend Next.js, React, TypeScript
Styling Tailwind CSS v4
Agents Strands Agents SDK
Model provider Amazon Bedrock
Reasoning model Amazon Nova 2 Lite
Voice model Amazon Nova 2 Sonic
Whiteboard tldraw
Realtime transport Socket.IO
Math execution mathjs, nerdamer, Node vm sandbox
Document handling pdfjs-dist

Project Structure

src/
  agents/          Strands agent definitions
  app/             Next.js app router and API route
  components/      UI modes and workspace components
  lib/             shared client/server helpers
  tools/           agent tool definitions
  types/           shared TypeScript types
server.ts          voice server for Nova Sonic
voice.sh           helper script for voice mode

Running Locally

Install dependencies:

pnpm install

Start the app:

pnpm dev

Start voice mode server in a second terminal:

pnpm dev:voice

Optional AWS login helper:

./scripts/login.sh

Why This Project Exists

Students do not study in one modality. They read notes, inspect diagrams, solve equations, revise with flashcards, and talk through ideas out loud.

Parchment brings those study behaviors into one agentic workspace powered by Amazon Nova.

About

for nova ai hackathon

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages