FDA & Trial Signal Detector

Vendors that sell into biotech — clinical research organizations, patient recruitment agencies, regulatory consultants — all have the same problem: they don't know a company needs them until that company is already talking to someone else. I built a tool that watches the public record instead — FDA filings, clinical trial registrations, grant announcements — for five specific events that predict a company is about to need outside help, and ranks which ones need it most urgently. Nobody asked me to build this one. I wanted a real signal-detection system to reason through, not a toy.

The first version got something wrong in a way that mattered: it was flagging academic hospitals and individual investigators as “inexperienced,” which is a different thing from “doesn't need a vendor.” A hospital running fifty clinical trials isn't inexperienced — it just runs trials in-house, so it's a bad sales target no matter what the raw signal looks like. I caught it, then fixed it two ways in the same night: a straight database patch for the clear cases, and a better prompt with explicit sponsor-type detection for the ones that actually needed judgment. Total cost: about two dollars. Classification accuracy on the hardest category went from roughly a coin flip to 75-85 percent.

I didn't chase it to perfect, either. The output is explicitly tiered — some signals are good enough to act on today, some need a second look, some should be skipped entirely — because treating a lead list as uniformly reliable is its own kind of mistake. Running today: 900+ signals across 700+ companies.