Sales Intelligence Agent
Juadsuarezsan/sales-intelligence-agent
Pesquisa B2B automatizada com planner-executor-reflector para prospecção em escala
Ver no GitHubAgent autônomo que pesquisa empresas B2B, constrói perfis estruturados e gera emails de prospecção personalizados. Implementa o padrão plan-execute-reflect usando Tavily, HackerNews e fetch web, produzindo CompanyProfile e OutreachEmail com sinais contextualizados.
Instalação
npx convez-ai Juadsuarezsan/sales-intelligence-agent1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Geração automática de dados de empresas para prospecção B2B
- Criação de emails personalizados com sinais recentes e pain points
- Automação do pipeline de vendas em sales-tech (Apollo, Clay, Outreach)
Veredito

Projeto bem arquitetado com padrão industrial comprovado (plan-execute-reflect via LangGraph). Ainda em fase planned, mas o roadmap é claro e ambicioso, ideal para quem quer entender automação de prospecção em escala com LLMs.
README
A tradução para o português deste README ainda está na fila.
Project 03 — B2B Sales Intelligence Agent
Planner-executor-reflector agent that researches target companies on the public web, builds structured profiles, and writes personalized outreach. The pattern behind Apollo.io, Clay.com, Outreach.io.
Industrial use case: sales-tech (Apollo, Clay, Outreach). AI Engineers who can ship this are ultra-demanded.
What this project does
Given a list of target companies, the system loops: plan what to research → execute searches (Tavily + HN + website fetch) → reflect on quality of info → refine. Produces a structured CompanyProfile and a personalized outreach email with measurable personalization.
Architecture
For each target company:
│
▼
[Planner] Claude → research plan
│
▼
[Executor loop] LangGraph
│ Tavily search · HN search · Website fetch
│
▼
[Reflector] Claude → "enough info? contradictions?"
│ if not enough → refine queries, loop
│
▼
[Profile Builder] Claude + Pydantic → CompanyProfile
│
▼
[Personalization Extractor] → pain points + recent signals
│
▼
[Email Writer] Claude → subject, hook, value prop, CTA
│
▼
[Quality Scorer] LLM-as-judge → personalization, accuracy, CTA clarity
Roadmap to v1.0.0
- Load YC companies dataset (~5K with metadata)
- Tavily API integration with retries + rate limiting
- selectolax-based website fetcher (faster than BeautifulSoup)
- LangGraph plan-execute-reflect loop with configurable max iterations
- Pydantic schemas for
CompanyProfile,OutreachEmail - Quality scorer using Claude Opus as LLM-as-judge with explicit rubric
- Eval set: 200 YC companies with ground truth metadata
- Batch processing of 100 YC companies, results persisted
- Next.js gallery view of the 100 pre-processed companies
- LangSmith public trace gallery with ≥30 agent loops
Stack
| Layer | Technology |
|---|---|
| LLM | Claude Sonnet 4.5 |
| Web search | Tavily API ($0.001/search, 1K free/mo) |
| News | HackerNews Algolia (free) |
| HTML parsing | httpx + selectolax |
| Structured outputs | Pydantic v2 |
| Orchestration | LangGraph (agent loop) |
| State | PostgreSQL |
| Frontend | Next.js 14 gallery view |
| Observability | LangSmith |
Definition of Done — project-specific
- Batch processing of 100 YC companies completed with persistent outputs
- Demo gallery shows all 100 pre-processed companies
- Comparison against 2 baselines: template-only + single-search+email, with numbers
- LangSmith trace gallery with ≥30 agent loops linked
- Cost vs quality analysis documented: at what point does more reflection stop improving
Plus the 12 universal DoD blocks.
License
MIT.