Professional work · Research system · AI

Window-profile research

A staged pipeline that researches window and door manufacturers and prepares profile-system data for WinFactor onboarding.

A browser-driving language-model agent reads each manufacturer site and returns the profile-system data as structured JSON.

01

Research pipeline

  • The pipeline scrapes manufacturer websites, refines systems into product lines, and assembles per-country JSON.
  • Its source workbook lists 1,426 manufacturers across 43 countries.
  • Each brand gets its own agent session. The agent drives a browser through an MCP server and extracts the fields of the profile-system schema.
  • A validation step checks the returned JSON and resumes the same session to repair a rejected or malformed answer.

02

Operation

  • Docker workers run on two remote machines, with Tailscale for connectivity and a local coordination machine.
  • The output follows the schema used by the WinFactor profile-system onboarding flow.

03

Technology

  • Python
  • Browser agents
  • MCP
  • JSON
  • Docker
  • Tailscale