We Killed Our Best Feature (the Numbers Made Us)

Every company tells you about the features they shipped. This is about the one we killed.

While building Geneziz AI — the local intelligence that organizes your library — we also trained something extra: a voice for the app. We tuned it on our own hardware, measured it against every conversation result we'd ever logged, and it was the best we'd ever had. Team morale said ship it.

Our quality gate said no. It cost the organizer about nine points of accuracy, and the organizer is the product. A rule we wrote before the experiment — ship nothing that makes the core worse — decided. So the voice went into the docs as a future opt-in, not into the app as a checkbox.

The gate was written before the experiment

That's the part people skip. We didn't debate after seeing the numbers; we registered the pass/fail rule first, then ran the training. When the result landed below the bar, there was nothing to argue about. We also tried stacking both roles in one model — same trade, same verdict. Two worlds, denied by measurement.

The same discipline caught something embarrassing: an old performance number we had published (72ms) was attributed to the wrong setup. We fixed it publicly and re-published the honest table: the real numbers are better anyway.

The payoff: 16× faster, on a CPU

Rethinking how the organizer reads context made everything dramatically faster: what used to take about 8.5 minutes for a thousand favorites now takes about 32 seconds — on a plain CPU, fully offline, for free. With a GPU, even faster.

And here's the discovery that shaped the whole product: your personal data is signal to you and noise to everyone else. We measured it — generic training data plateaus no matter how much you add, and strangers' data adds nothing measurable. So the architecture is honest: one small, well-trained universal model that ships inside the app (~1.6 GB), plus a tiny personal layer that learns from your confirms — on your machine, in seconds, never leaving it.

Trained at home, on our own machines

No rented GPUs. No datacenter bill. The training ran at night, at home, on a gamer-grade graphics card we already owned — every generation about 55 minutes. When the numbers said a technique was hurting the model, we froze the recipe and never looked back.

That's what "local-first with AI" means to us: zero cloud cost, zero data leaving the house, and a quality bar that can say no to our own favorite ideas.

Geneziz: geneziz.app — the intelligence that lives in your machine, and the honesty to keep it that way.