Lead Gen: AI lead research

Tell it what you sell. Get a list of local companies likely to buy it.

Try the demo (no sign-up, nothing to install)

For each company you get:

• what they do, in one sentence
• a fit score from 0 to 100, with the reason
• a draft opening line for a cold email
• whether they're already in your CRM

An example

Searching for customers for EV charge points, the top lead was a vehicle hire firm:

Score: 92. A vehicle hire firm with two branches, running a fleet of cars and vans and offering MOT and servicing.

Opener: "Your Blyth and Bedlington branches run a self-drive car and van fleet alongside MOT and servicing work, so vehicles are moving through your sites every day."

How it works

flowchart LR
    A[Find companies] --> B[Find their websites]
    B --> C[Read the websites]
    C --> D[AI profiles and scores each one]
    D --> E[Check against your CRM]

1. Find companies from Companies House, the official UK register. This demo covers small companies in the North East of England.
2. Find their websites with a web search, and check each site really belongs to the company.
3. Read each website: the homepage and about page.
4. Profile and score each company with an AI model (Claude), against a plain-English description of your ideal customer.
5. Check against your CRM, so you only get companies you don't already have. New leads are exported as a file ready to import into HubSpot.

What the demo run found

• Of 200 small companies sampled, 55 had a website that was verifiably theirs. Most small companies don't have one.
• 35 were new leads; the other 20 were already in the (synthetic) CRM.
• The AI caught wrong website matches that simple rules missed, such as a US company with a similar name.
• AI costs are about $0.54 per 1,000 companies profiled.

Data and privacy

• Company-level public data only: Companies House records and each company's own public website.
• No personal data: no names, emails or phone numbers of individuals, and no LinkedIn. Emails and phone numbers are removed from website text before it is stored. This keeps it simple under UK GDPR.
• Polite scraping: it respects robots.txt, waits between requests and identifies itself.
• Synthetic CRM: the CRM data used for matching is made up.

---

For developers

Built with Python, httpx, BeautifulSoup, DuckDB, Pydantic, Claude Haiku 5.5 and Next.js.

You need a Companies House API key, a Serper.dev key (web search) and an Anthropic API key.

conda create -n leadgen python=3.12
conda activate leadgen
pip install -r requirements.txt
cp .env.example .env   # then add your API keys

Each stage runs on its own and can be re-run safely; it only processes companies that don't have a result yet.

python -m lead_gen.db                # create the database
python -m lead_gen.find_companies    # 1. find companies
python -m lead_gen.find_websites     # 2. find websites
python -m lead_gen.scrape            # 3. read the websites
python -m lead_gen.enrich            # 4a. AI profile
python -m lead_gen.score             # 4b. AI fit score and opener
python -m lead_gen.synthetic_crm     # 5a. generate the synthetic CRM export
python -m lead_gen.crm_clean         # 5b. clean, de-duplicate and match
python -m lead_gen.hubspot_export    # 5c. HubSpot-ready CSV of new leads
python -m lead_gen.export_dashboard  # data snapshot for the web demo

To run the web demo locally, see web/README.md.

• The model never browses or searches. It only reads text the pipeline has already collected.
• Every answer must match a fixed Pydantic schema before it is stored.
• It is told to use only what the website says, and to leave a field empty rather than guess.
• Scoring reads the short profile, not the whole website, which keeps it cheap.
• Every call is cached and logged with its token count and cost, and prompts are versioned.
• Answers were checked against 25 hand-labelled companies: see eval/results.md.

lead_gen/          one Python module per pipeline stage, plus config, db and shared helpers
web/               the Next.js demo site
eval/              labelling sheet and evaluation reports
docs/findings.md   notes and numbers from every stage

The customer descriptions used for scoring are in CUSTOMER_DESCRIPTIONS in lead_gen/config.py.