Sales
Lead Research with LLM Extraction
Challenge
A sales organisation researched potential customers such as offices and law firms by hand: search addresses, open websites, pick out contact persons, check against existing customers, transfer to Excel. Hours per region, with patchy results.
Solution
We developed a Windows desktop app that searches for offices and law firms via the Google Places API, scans company websites for contact persons (rule-based or LLM-assisted), matches the hits against the existing customer base and exports the result to Excel. Everything runs locally on the machine, with no server.
Result
Hours of manual work become a single search run with a finished Excel list. Because no data leaves the device, there were no hurdles with data protection or IT approval.
Starting point
New customer acquisition starts with a list. At the customer, that list was created manually: an employee searched a region for offices and law firms, opened every website, looked for the right contact person and checked whether the company was already a customer. The result was slow, inconsistent and hard to repeat when the next region came up.
What the app does
The user enters a location, radius and industry. The app queries the Google Places API and returns the hits with address, phone and website. It then visits the websites and extracts contact persons. There are two modes for this: a rule-based mode that scans legal notice and team pages for known patterns, and an LLM-assisted mode that also understands unstructured pages and cleanly pulls out names, roles and contact details.
Before the list is exported, the app matches every hit against the existing customer base. Known companies are flagged so that sales does not contact existing customers again. The export is an Excel file in the familiar format.
Why local and without a server
Contact data is sensitive, and the customer did not want a cloud platform for its lead lists. The app therefore runs entirely on the user’s Windows machine. There is no server, no database on the network and no dependency on a web application. This also simplified the IT approval process: one executable, one API key, done.
Technology
The application is written in Python with the interface in Qt/QML. PyInstaller turns it into a single Windows application with no installation effort. The LLM mode can be switched on optionally, so the user decides when the model’s accuracy justifies the effort.
Transferability
The pattern fits any research that turns public sources into structured lists: suppliers, partners, trade fair exhibitors, dealers. Search, enrich, match, export.