SALES ICE-BREAKER TOOL - DISCUSSION NOTE

Topic: Internal tool to help sales agents ice-break with prospects on first contact.

Idea: Pull public, profession-related info about a prospect from multiple sources and produce a short brief with 3-5 fact-grounded talking points the agent can use.

Input: Email, or name + company, or LinkedIn URL.
Output: One-page brief in under 30 seconds - career snapshot, recent activity, achievements, topics, and cited ice-breakers.

Sources: LinkedIn (via Proxycurl), GitHub, Twitter/X, Google News, podcasts, YouTube, personal sites/blogs, email-enrichment APIs.

Approach:
- Identity Resolver -> Orchestrator -> parallel source Connectors -> Aggregator -> LLM Summarizer (Claude) -> Renderer.
- Cache layer with per-source TTL to avoid re-billing APIs on repeat lookups.
- Every LLM-generated claim must cite a source - no hallucinations.

Stack: Python 3.12, FastAPI, asyncio, SQLite (MVP), Claude Sonnet 4.6, Streamlit for UI.

Phasing:
- Phase 1 MVP (~1-2 wks): CLI + LinkedIn + GitHub + News, LLM brief, file cache.
- Phase 2 (~2-3 wks): more connectors, web UI, audit log, email-only entrypoint.
- Phase 3 (~1-2 mo): CRM integrations, scheduled refresh, multi-tenant.

Cost: ~$0.20 per cold-cache brief; ~$0.02 warm.

Constraints:
- No direct LinkedIn scraping - must use a licensed proxy (Proxycurl).
- Public professional info only; GDPR/CCPA-aware retention.
- Citations mandatory on every claim.

Non-goals: Lead scoring / ICP fit, drafting the outreach message itself, anything behind a login wall.

Next steps:
1. Confirm scope and non-goals.
2. Pick project location, initialize Python repo.
3. Get API keys: Proxycurl, Apollo, NewsAPI, Anthropic.
4. Build Phase 1 vertical slice end-to-end.
5. Translate architecture into a draw.io diagram.

Open questions:
- Buy vs build for LinkedIn data (leaning Proxycurl on legal + effort grounds).
- Single-agent tool or org-wide deployment?
- Should it eventually draft the outreach message too? (Out of scope for now.)
