The short version

QSO Predictor is written by me — Peter Hirst, WU2C, in grid FN42 (northeastern Massachusetts). First licensed in 2022, Amateur Extra, modest station: wire antenna and a 600-watt tube amplifier. Nearly everything the app does exists because something puzzled or bit me at my own station — the rest came from sharp-eyed early users — and everything ships only after it has worked on the air here.

The longer version

As a teenager I was a shortwave listener. A license — and the equipment to do more than listen — both seemed like a long shot. Life went another way for a while: a degree and PhD in physics, working on electronics and the physics of high-power microwave devices, including some experimenting with early neural-network modelling in the early 1990s. Many years later I ended up at MIT, working in management education.

During the COVID lockdown I discovered that a ham license was within far easier reach than the teenage SWL remembered — and took my exams in person at MIT, courtesy of the MIT Radio Society, W1MX. Licensed at last — as KC1RFZ, soon WU2C — and now with a few resources that teenager never had. Not unlimited ones: a wire antenna and an old-school tube amp will only take you so far on SSB, though a few true DX contacts made it through.

Then one day I installed WSJT-X, and FT8 was a revelation: real DX, modest station, physics doing the heavy lifting.

But one thing kept nagging me. I’d pick an apparently empty offset, call a DX station, and… nothing. Again and again — even when PSK Reporter showed my signal was being decoded in the DX station’s own region. The band map said the frequency was clear. The propagation said I was getting there. So why no answer?

Anyone who has used QSO Predictor knows how that story ends: the frequency was only clear on my end. The competition — the pileup — was often invisible to me, and the answers were scattered across data sources nobody had assembled into one live picture.

The code factory

Over Thanksgiving 2025 I started exploring AI coding tools, and the idea of a digital-modes helper became the test project. It stopped being a test very quickly. Working with a variety of AI tools — eventually converging on Claude Code — turned into what I can only describe as a one-human code factory: I bring the operating problems, the radio physics, the design judgment, and the on-air testing; the AI brings velocity. There’s a pleasing symmetry in it, thirty years after those first neural-net experiments.

Every feature is specified, reviewed, and field-tested by me at this station before it ships — usually because I needed it the night before. The hidden-pileup display, the Audio Doctor, the station checkup, click-to-call: most of them started as my own problems.

But not all of them, and not only mine. A small group of early users has shaped this app more than its size suggests — reporting the bugs I couldn’t reproduce, suggesting the fixes I wouldn’t have found, and testing on stations and setups I don’t have. Brian KB1OPD, Warren KC0GU, W6IX, and others are credited by name in the release notes and commit history where their reports and ideas landed. A one-human code factory still runs better with good field reports.

QSO Predictor is free, open source (GPLv3), and collects no data of any kind. It never transmits on its own. It exists because I wanted it, and it’s shared because the puzzle it solves isn’t unique to me.

73, Peter WU2C

Questions, bug reports, or your own hidden-pileup story: GitHub or see Support.