buyer-agent
WiseGitter/buyer-agent
A small Python pipeline that uses Claude, the Wikipedia/Wikidata APIs, and a SQLite event log to discover, score, enrich, and draft outreach for European B2B sales leads.
Instalação
npx convez-ai WiseGitter/buyer-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
README
A tradução para o português deste README ainda está na fila.
Buyer Agent
A small autonomous pipeline that discovers, scores, and enriches European B2B sales leads — built originally to find truck and trailer manufacturers in Germany and Poland for an import-export business from India.
The interesting part isn't the LLM; it's the scaffolding around it. This repo is a worked example of how to wrap an LLM into a real, observable, idempotent system.
What it does
A single python main.py runs the full pipeline:
- Discover candidate manufacturers via the Wikipedia and Wikidata APIs
(Wikidata's
P856property gives curated official-website URLs — far more reliable than scraping Wikipedia HTML). - Score each company against a product catalog using Claude (Haiku for cost-effective scoring; Sonnet for higher-quality drafting). The scoring step uses forced tool-use to guarantee structured output — no JSON parsing, no markdown fences.
- Enrich each scored company by crawling its own website for published procurement emails. Multilingual keyword detection (EN / DE / PL) finds contact, imprint, and supplier pages. A heuristic ranker picks the best email and down-weights GDPR / DPO addresses.
- Draft personalized outreach for the strongest matches, with the scoring step's reasoning piped forward as the email's hook.
- Sync to a Notion CRM via the Notion REST API. Subsequent runs update existing pages (create-or-update / upsert pattern) rather than creating duplicates.
A human approval gate in Notion sits between drafted outreach and anything that actually reaches a real inbox. Nothing sends automatically.
Architecture
┌──────────┐ ┌──────┐ ┌────────┐ ┌──────┐ ┌────────┐
│ Discover │──▶│Score │──▶│Enrich │──▶│Draft │──▶│ Sync │
│ │ │ │ │contact │ │ │ │ Notion │
└──────────┘ └──────┘ └────────┘ └──────┘ └────────┘
│ │ │ │ │
└────────────┴───────────┴───────────┴───────────┘
│
▼
┌──────────────────────────┐
│ SQLite source of truth │
│ (state + event log) │
└──────────────────────────┘
│
▼
┌──────────────────────────┐
│ Streamlit dashboard │
│ (analytical surface) │
└──────────────────────────┘
- SQLite is the canonical state. Two tables:
companies(mutable per-row state) andevents(immutable append-only log of every API call — model, tokens, cost in USD, latency, errors). - Notion is the operational surface — where you work the leads.
- Streamlit is the analytical surface — KPI cards, funnel chart, fit-score histogram, cumulative cost over time, recent activity.
- The pipeline is idempotent — re-runs pick up where they left off, never re-process, never double-charge.
Stack
Python · Anthropic API (Claude Haiku + Sonnet, forced tool-use) · Wikipedia + Wikidata APIs · Notion API · BeautifulSoup · SQLite · Streamlit · Tenacity (retries with exponential backoff).
Sample result
First run on 26 manufacturers in Germany and Poland:
- Working procurement emails extracted for ~75% of companies.
- Failures were almost all JS-rendered sites that would need a headless browser to crawl.
- Total LLM cost: under $0.20.
Run it yourself
Heads up: running this uses your own Anthropic and Notion API keys. Expect roughly $0.20 per 26-company batch. The repo ships with no credentials.
- Clone the repo and create a Python virtualenv.
pip install -r requirements.txt- Get an Anthropic API key and a Notion integration token.
- Create a Notion database with the right schema (see
buyer_agent/storage.pyfor the expected property names) and share it with your integration. - Create a
.envfile at the project root:
ANTHROPIC_API_KEY=sk-ant-...
NOTION_API_KEY=ntn_...
- Update
DATA_SOURCE_IDinconfig.pywith your Notion data-source ID. - Create your product catalog by copying the example:
cp buyer_agent/catalog.py.example buyer_agent/catalog.py
Then edit buyer_agent/catalog.py to describe what your business sells.
The catalog is the reference frame Claude uses to score companies.
8. Run: python main.py
9. View the dashboard: streamlit run dashboard.py
Project layout
.
├── main.py # orchestrator + CLI
├── dashboard.py # Streamlit analytics dashboard
├── config.py # settings + secrets loader
├── requirements.txt
└── buyer_agent/
├── catalog.py # product catalog (what we sell)
├── discovery.py # candidate-discovery interface
├── scoring.py # Claude scoring (forced tool-use)
├── drafting.py # Claude outreach drafting
├── storage.py # Notion REST API (create + update)
├── database.py # SQLite layer (schema as code)
└── scrapers/
├── wikipedia.py # Wikipedia + Wikidata discovery
└── contacts.py # company-website email scraper
What's next
- Playwright-based scraper for the IAA Transportation exhibitor directory (JavaScript-rendered SPA — needs a headless browser).
- Second enrichment pass using the LLM to infer company size and procurement structure from the Wikipedia summary.
- Scheduled runs (cron / GitHub Actions) so discovery picks up newly-published companies over time.
License
MIT — see LICENSE.