I Built an AI That Does 30 Minutes of Sales Research in 90 Seconds. Here's How Nova Made It Possible.

Every sales professional knows the drill. Before a call, you open five browser tabs: the company website, LinkedIn, Crunchbase, Google News, and their careers page. Thirty minutes later, you have scattered notes and your meeting is about to start.

Scout eliminates this entire workflow.

Sales teams spend 20-30% of their week on pre-call research. It's the same process every time: visit the same five websites, extract the same categories of information, and synthesize it into something useful. This is exactly the kind of task AI should handle.

Scout is an AI-powered company research agent that takes a company name and returns a complete briefing in under two minutes. What makes it unique is how deeply it integrates Amazon Nova services.

Nova Act is Scout's eyes. Instead of relying on stale training data, Nova Act navigates real websites in real time. It visits the company homepage, searches Google News, browses LinkedIn, checks Crunchbase, and scans job listings. This isn't web scraping — it's intelligent navigation that understands page structure and extracts relevant data.

Raw extracted data isn't useful on its own. Nova 2 Lite via AWS Bedrock acts as Scout's brain, synthesizing findings from all five sources into a structured briefing. It identifies key people, summarizes recent news, confirms the tech stack from job listings, detects growth signals from hiring patterns, and generates conversation starters that reference specific findings.

The synthesis prompt instructs Nova 2 Lite to act as a sales intelligence analyst, cross-referencing data across sources and flagging inconsistencies. A confidence score reflects how many sources returned useful data.

This is where Scout becomes more than a one-shot tool. After every research job, Scout generates a 384-dimensional embedding vector using Nova Multimodal Embeddings via Bedrock's InvokeModel API. These embeddings capture the semantic meaning of each briefing.

The result: semantic search across all past research. A sales manager can search "AI safety company" and Scout returns Anthropic at the top — not because the word "Anthropic" appears in the query, but because the embedding space understands what these companies actually do. Search "payment processing fintech" and Stripe surfaces first.

This transforms Scout from a research tool into a knowledge base that grows smarter with every query.

For individual sales reps, Scout saves 30 minutes per call. For a team of 50 making 10 calls per day, that's 250 hours per week — over $500,000 per year in recovered productivity at median sales salaries.

But the real impact is qualitative. Better-prepared reps have better conversations. Talking points that reference a company's recent AI fraud detection launch or their expansion into African markets signal genuine interest. This builds trust, shortens sales cycles, and improves close rates.

Scout is designed for B2B sales teams, account executives, business development representatives, and anyone who needs to understand a company quickly. The semantic search feature makes it particularly valuable for sales organizations that research hundreds of companies per quarter — previous research is never wasted.

Scout is open source on GitHub. The architecture is modular: teams can add new data sources (SEC filings, patent databases, social media) without changing the synthesis layer. The embedding-based search means new sources automatically improve search quality.

The roadmap includes team collaboration features (shared research libraries), CRM integrations (auto-populate Salesforce/HubSpot), and scheduled research (monitor companies over time with automatic alerts on changes).

Scout runs on FastAPI (Python backend) and Next.js (frontend). The backend orchestrates five extractors in sequence, pushes real-time progress via Server-Sent Events, and stores everything in SQLite. No external vector database is needed — Nova Multimodal Embeddings produces normalized vectors that work with simple cosine similarity in pure Python.

The entire system degrades gracefully: if a source is blocked, Scout produces a briefing from whatever succeeds. If embeddings fail, search returns empty results instead of crashing. This resilience makes Scout production-ready from day one.