A multi-agent AI research paper analysis tool built for the Ignite USD Hackathon.

Upload a PDF or paste a link to any academic paper. ScholarAI runs multiple AI agents in parallel to break it down from every angle: summary, deep explanation, code extraction, future research directions, and interactive chat. All personalized to your expertise level and goals.

The agents are model-agnostic. Each agent slot can call whichever AI model is best suited for that category. The current implementation uses Anthropic Claude (Haiku for speed-critical tasks, Sonnet for depth), but any specialized model can be swapped in per agent.

• Harshith Reddy Nalla - Portfolio
• Smriti Regmi - LinkedIn

1. Upload a research paper (PDF drag-and-drop) or paste any URL (ArXiv, etc.)
2. Fill a short questionnaire to personalize the output to your expertise, goals, and preferences
3. Get analysis across 5 tabs, all running simultaneously:
   • Summary: title, authors, problem, methodology, results, datasets with links, keywords
   • Deep Explainer: section-by-section breakdown, KaTeX-rendered equations, SVG diagrams
   • Code: extracted algorithms with syntax highlighting, simplified versions, setup guides
   • Future Research: limitations, open questions, research ideas, big picture
   • Chat with Paper: full streaming chat powered by the complete paper text
4. Every tab has a "Dive Deeper" bar at the bottom for context-aware follow-up questions without leaving the panel

Every agent's system prompt is prepended with the user's profile so the AI adjusts vocabulary, depth, analogy choice, and format automatically.

Each agent is an independent module that can call any AI model. The current configuration:

After the summary returns, three agents fire simultaneously without waiting for each other:

Summary done
  |-- Deep Explainer   (analysis + diagrams in parallel)
  |-- Code Explainer
  |-- Future Research

Before analysis runs, the user fills a questionnaire covering expertise level, goal, pain points, format preference, and optional free-text instructions. This gets converted into a reader profile block that is prepended to every agent's system prompt.

The frontend never calls Anthropic directly. All requests go through Express endpoints:

• POST /api/claude for non-streaming calls (returns full JSON)
• POST /api/claude/stream for streaming calls (pipes SSE to client)
• GET /api/fetch-url for fetching paper content from URLs

The API key lives only on the server. The backend validates a model allowlist and enforces token limits per request.

An AbortController is created when analysis starts. Every agent call receives its signal. Clicking "New Paper" or navigating away aborts all in-flight requests instantly.

ScholarAI/
├── server.js                        Express backend, proxies all AI calls
├── .env                             ANTHROPIC_API_KEY + PORT (gitignored)
├── vite.config.ts                   Vite config with /api proxy
└── src/
    ├── App.tsx                      Root component, state management, agent orchestration
    ├── main.tsx                     Entry point
    ├── index.css                    Global styles, scrollbar, fonts
    ├── types/
    │   └── index.ts                 All TypeScript interfaces
    ├── agents/
    │   ├── summarizerAgent.js       Structured JSON summary
    │   ├── deepExplainerAgent.js    Deep analysis + SVG diagrams
    │   ├── codeExplainerAgent.js    Code extraction + explanations
    │   ├── futureResearchAgent.js   Limitations + research ideas
    │   └── chatAgent.js             Streaming chat wrapper
    ├── components/
    │   ├── HomePage.tsx             Landing page with video background
    │   ├── Questionnaire.tsx        Personalization modal
    │   ├── ResultsPage.tsx          Main results layout with tab routing
    │   ├── TabBar.tsx               Pill-style navigation bar
    │   ├── AppShell.tsx             Root layout wrapper
    │   ├── DropZone.tsx             PDF drag-and-drop upload
    │   ├── panels/
    │   │   ├── SummaryPanel.tsx
    │   │   ├── DeepExplainerPanel.tsx
    │   │   ├── CodePanel.tsx
    │   │   ├── FuturePanel.tsx
    │   │   └── ChatPanel.tsx
    │   └── ui/
    │       ├── AnimatedCard.tsx     Fade-in card wrapper
    │       ├── CodeBlock.tsx        Syntax-highlighted code with copy
    │       ├── EquationBlock.tsx    KaTeX math rendering
    │       ├── Prose.tsx            Markdown renderer
    │       ├── KeywordPill.tsx      Keyword tags
    │       ├── SectionDivider.tsx   Visual separator
    │       ├── GlowCard.tsx         Glowing card variant
    │       ├── LoadingBar.tsx       Progress indicator
    │       └── SuggestedQuestions.tsx  Chat starter prompts
    ├── utils/
    │   ├── api.js                   callClaude() + streamClaude()
    │   ├── extractText.js           PDF + URL text extraction
    │   ├── buildUserProfile.js      Questionnaire answers to prompt string
    │   └── formattingRules.js       Shared formatting rules for agents
    └── hooks/
        └── usePdfExtractor.ts       PDF file text extraction

• Node.js 18+
• An Anthropic API key

git clone https://github.com/HarshithReddy01/ScholarAI.git
cd ScholarAI
npm install

Create a .env file in the project root:

ANTHROPIC_API_KEY=sk-ant-your-key-here
PORT=3001

npm run dev

This starts the Express backend (port 3001) and Vite dev server (port 5173) concurrently. Open http://localhost:5173.

Other commands:

npm run dev:server   # backend only
npm run dev:client   # Vite only
npm run build        # production build
npm start            # serve production build

• API key lives only in .env on the server, never sent to the browser
• SVG diagrams from AI are sanitized before rendering (strips scripts, event handlers, javascript: hrefs)
• Backend validates a model allowlist and enforces token caps
• .env is gitignored