Scout - AI Company Research Agent

Built for the Amazon Nova AI Hackathon 2026

Hackathon Submissions

Hackathon	Platform	Track	Status
Amazon Nova AI Hackathon 2026	DevPost	Open	SUBMITTED (no Nova Act key -- geo-blocked, HTTP fallback used)

Live Demo & Links

Live Frontend	frontend-murex-eta-95.vercel.app
Live Backend API	scout-api.astraedus.dev (API docs)
Demo Video	youtu.be/t9kPFHv62m4
Blog Post	How Scout Uses Three Amazon Nova Services
DevPost	devpost.com/software/scout-ixslhg

Special adaptations: Added http_website.py + http_news.py extractors so Scout works without Nova Act API key (geo-blocked outside US). Three modes: mock, http-fallback (requests/BS4 + Bedrock), nova-act.

Scout turns company research from a 30-minute manual task into a 2-minute automated briefing. Type a company name, and Scout's AI agents navigate real websites to gather live data, then synthesize it into actionable intelligence for sales calls and meetings.

The Problem

Before every sales call or meeting, professionals spend 30-60 minutes manually browsing:

• Company website for products and team info
• LinkedIn for employee count and key people
• Crunchbase for funding history
• Google News for recent developments
• Job boards for growth signals and tech stack

The Solution

Scout automates this entire workflow:

1. Input a company name
2. Nova Act navigates 5+ real websites, extracting structured data from each
3. Nova 2 Lite synthesizes all findings into a structured briefing
4. Dashboard displays the briefing with evidence and talking points

Key Features

• Multi-site research: Extracts data from company websites, LinkedIn, Crunchbase, Google News, and careers pages
• Live data: Nova Act browses real websites in real-time (not stale training data)
• Structured briefings: Key people, recent news, tech stack, growth signals, competitive landscape, and suggested talking points
• Graceful degradation: If a source is blocked or unavailable, Scout still produces a briefing from whatever succeeds
• Real-time progress: SSE-powered progress tracking shows each research stage live
• Research history: SQLite-backed history of past research jobs

Architecture

graph LR
    A[Next.js Frontend<br/>Search + Dashboard] -->|Search query| B[FastAPI Backend<br/>EC2]
    B -->|Extract listings| C[Amazon Nova Act<br/>Web Scraping Agent]
    C -->|Structured listing data| B
    B -->|Analyze + Enrich| D[Nova Micro<br/>Lead Scoring]
    D -->|Scored leads| B
    B -->|Lead cards +<br/>Contact info| A

Loading

Tech Stack

Layer	Technology
Frontend	Next.js 14, TypeScript, Tailwind CSS
Backend	Python, FastAPI, Uvicorn
Browser AI	Amazon Nova Act (website navigation + extraction)
Synthesis AI	Amazon Nova 2 Lite via AWS Bedrock (reasoning + structuring)
Semantic Search	Amazon Nova Multimodal Embeddings via Bedrock (384-dim vectors)
Database	SQLite via aiosqlite
Real-time	Server-Sent Events (SSE)

Getting Started

Prerequisites

• Python 3.10+
• Node.js 18+
• AWS account with Bedrock access (Nova 2 Lite)
• Nova Act API key from nova.amazon.com/act

Backend Setup

# Create virtual environment
python3 -m venv venv
source venv/bin/activate

# Install dependencies
pip install -r backend/requirements.txt

# Configure environment
cp .env.template .env
# Edit .env with your API keys

# Run (mock mode for testing without API keys)
MOCK_MODE=true uvicorn backend.main:app --reload --port 8000

# Run (production mode with real APIs)
MOCK_MODE=false uvicorn backend.main:app --port 8000

Frontend Setup

cd frontend
npm install
npm run dev

Visit http://localhost:3000

Test API Connections

source venv/bin/activate
python scripts/hello_nova_act.py    # Test Nova Act
python scripts/hello_bedrock.py     # Test Bedrock

API

Method	Endpoint	Description
POST	/api/research	Start research (body: {"company_name": "Stripe"})
GET	/api/research/{id}	Get status and results
GET	/api/research/{id}/stream	SSE progress stream
GET	/api/history	Recent research history
GET	/api/search?q=	Semantic search across briefings (Nova Embed)
GET	/health	Health check

How It Uses Amazon Nova

Nova Act (Browser Automation)

Each extractor uses Nova Act to navigate real websites:

• act() for all browser actions — navigation, clicking, and data extraction
• Multiple browser sessions extract data from different sources in sequence
• Graceful error handling for blocked/unavailable sites

Nova 2 Lite (AI Synthesis)

After extraction, all data is fed to Nova 2 Lite via AWS Bedrock's Converse API:

• Synthesizes raw data into structured JSON briefing
• Generates talking points referencing specific findings
• Assesses data quality and confidence level
• Identifies growth signals and competitive positioning

Nova Multimodal Embeddings (Semantic Search)

After synthesis, each briefing is embedded using Nova Multimodal Embeddings via Bedrock's InvokeModel API:

• Generates 384-dimensional embedding vectors for briefing content
• Enables semantic search across all past research ("find companies in AI safety")
• Uses GENERIC_INDEX purpose for indexing, TEXT_RETRIEVAL for queries
• Cosine similarity ranking returns the most relevant briefings
• Fully integrated: embeddings auto-generate after every research job

Project Structure

scout/
  backend/
    main.py              # FastAPI app + orchestration
    config.py            # Environment configuration
    extractors/          # Nova Act browser extractors
      website.py         # Company website extractor
      google_news.py     # Google News search
      linkedin.py        # LinkedIn company page
      crunchbase.py      # Crunchbase funding data
      careers.py         # Job listings extractor
      mock.py            # Mock extractors for development
    synthesis/
      briefing.py        # Nova 2 Lite synthesis
      embeddings.py      # Nova Multimodal Embeddings
      mock_briefing.py   # Mock synthesis for development
    models/
      schemas.py         # Pydantic data models
    db/
      database.py        # SQLite storage
  frontend/              # Next.js dashboard
  scripts/
    hello_nova_act.py    # Nova Act smoke test
    hello_bedrock.py     # Bedrock smoke test

Built By

Diven Rastdus - Full-Stack Developer & AI Engineer

#AmazonNova