ClientOS

The problem

Finding clients as a freelancer usually means hours of browsing Google Maps and websites, then sending messages that sound like everyone else's: "I noticed your website could be improved."

Using AI to speed this up creates a worse problem. A model will happily write that a site loads in 9 seconds or loses 40% of its visitors, and one invented number sent to a real business owner ends the conversation and damages the sender's name.

The hard part

Making every claim true. The model never supplies a number: it receives a block of measured data and may only cite its keys, a validator drops any finding that cites something else, and the email writer may describe how a problem costs customers but never attach a percentage it did not measure. The security checks follow the same rule. They only read what a site sends to any visitor, never probe hidden addresses, and stay silent in the edge cases, for example a PHP version number that a Linux distribution patches without changing it. Tested against 25 real sites, they raised no false alarm.

Where it stands

In daily use since September 2026 and now on its sixteenth version, deployed privately on Vercel. It started as ProspectIQ, a prospect finder, and grew into the whole loop: finding the business, proving the problem, the first message, the follow-ups, the proposal, the invoice and the case study. More than 480 automated tests cover the scoring, the guards on what the AI may claim, the mailbox limits and the reports.