SELINE INTEL — Sprint 1

AI-powered business intelligence research CLI. Sprint 1 scope per roadmap: Research Orchestrator + Website Agent + News Agent (search-fallback) + Report Generator.

Setup

npm install
cp .env.example .env   # optional: add SERPER_API_KEY for search fallback

Usage

npm run research -- "Hims & Hers"

Writes reports/<slug>.md.

Architecture

src/
  agents/
    orchestrator/   # coordinates agents, builds ResearchBundle
    website-agent/   # site discovery + heuristic extraction
    news-agent/       # search-based news collection
    report-agent/     # ResearchBundle -> markdown
  lib/
    providers.ts       # FetchProvider / SearchProvider abstraction
  types/
    research.ts         # shared data models (Source, CompanyProfile, etc.)
  cli.ts                  # commander entrypoint

Sourcing strategy

Each agent goes through resolveAndFetch(): try a direct URL guess first, fall back to web search (Serper.dev) if that fails or no key is set. No key -> degraded mode (direct fetch only), not a crash. This keeps agent logic decoupled from the provider, so swapping in a different search API or an LLM-based extractor later doesn't touch agent code.

Known Sprint 1 limitations (by design, not bugs)

• Website Agent extraction is heuristic (first-N-chars description), not LLM-parsed. Output shape is stable so this is a drop-in upgrade later.
• No Leadership/Funding extraction yet — needs structured parsing, planned for a later sprint.
• Risks/Opportunities sections are placeholders — that's analytical synthesis, not raw collection, and is intentionally out of scope here.
• Agents run sequentially, not in parallel (debuggability over speed for Sprint 1).

Roadmap

See project spec — Sprint 2 adds Corporate Agent + citation engine, Sprint 3 adds Supabase persistence, Sprint 4 the Next.js dashboard.