PaperPilot

An AI-assisted research workflow for searching arXiv, selecting papers, and previewing/synthesizing findings. Built with Next.js App Router, React, Nuqs for URL state, and a modern component system.

What it does

• Search arXiv and see rich paper cards with authors, categories, and links
• Select multiple papers and keep selection in the URL (shareable)
• Preview selected PDFs side-by-side with an AI Notes panel
• “Synthesize” mode (URL param mode=synth) for cross-paper insights

Why it’s useful for the AWS AI Agent Hackathon

• Demonstrates an end-to-end research agent UX: discovery → curation → analysis
• Clear integration points for AWS services:
  • Amazon Bedrock: model inference (Claude, Llama, etc.) for summarization/synthesis
  • Amazon S3: store uploaded PDFs or cached extractions
  • Amazon API Gateway + Lambda: lightweight inference proxy and orchestration
  • Amazon OpenSearch Serverless: semantic search or vector store for paper embeddings
  • Amazon Cognito: auth if you add user accounts and saved workspaces

This repo ships a polished front-end with URL-synced state via Nuqs. Hook in your AWS services to power the AI side.

Tech stack

• Next.js 15 (App Router), React 19
• Nuqs (URL query state)
• TanStack React Query
• Tailwind CSS + Radix UI primitives
• react-resizable-panels

Project structure

• app/ — routing, pages, layout, loading
• components/ — UI and feature components
• lib/ — utilities
• public/ — static assets

Local development

• Requirements: Node 18+, pnpm
• Install deps: pnpm install
• Run dev server: pnpm dev (http://localhost:3000)

Key flows

• Selection persists in the URL using two params:
  • sel: array of selected paper IDs
  • meta: URL-encoded JSON map of paper metadata keyed by ID
• Preview page (/preview):
  • Reads sel and meta via Nuqs
  • Renders PDFs via pdfLink and shows an AI panel
  • If mode=synth, the right panel copy changes to “Synthesis”

Where to connect AWS

• Create an API route (e.g., app/api/summarize/route.ts) that calls Bedrock
• Add a form/action in the Preview page’s right panel to POST selected meta
• Optionally store extracted text in S3 and index metadata in OpenSearch
• If doing agentic retrieval, add embeddings + vector search with Bedrock Titan or OpenSearch k-NN

Environment variables (example)

• BEDROCK_REGION — AWS region for Bedrock
• BEDROCK_MODEL — base model ID (e.g., anthropic.claude-3-5-sonnet-20240620-v1:0)
• AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY — if not using a role

Add a .env.local with these to enable server routes that call AWS.

Deployment

• Easiest path: Vercel (Next.js native)
• AWS path: Amplify Hosting or custom on ECS/Fargate + ALB
• Ensure environment variables are configured in your host

Submission checklist (Hackathon)

• Public repo with README and setup steps
• Deployed preview URL (Vercel or Amplify)
• Short demo video (2–4 min) showing search → select → preview → synth
• Note AWS services used and architecture diagram
• Instructions to reproduce (env vars, how to run)

Troubleshooting

• If Nuqs URL state doesn’t change, ensure NuqsAdapter wraps the app in app/layout.tsx
• If PDFs don’t render, verify pdfLink is accessible (CORS may block some hosts)
• If the UI looks off, confirm Tailwind is working and global styles are included

---

Made with ❤️ for the AWS AI Agent Global Hackathon.