Each candidate is an on-device AI with a hidden persona — genuinely brilliant, plausibly fraudulent, quietly burnt out, or brilliant-but-toxic. You interview them by voice, asking whatever you like. They answer in character. Then you hire, and three months later you find out what you actually got.
The score is how accurately you read people, not how many questions you asked.
What it does
Hold to talk and ask whatever you want. There is no question list and no branching script — the candidate answers in character.
Brilliant. Plausibly fraudulent. Quietly burnt out. Brilliant but toxic. You are not told which, and the good ones are hard to tell apart from the dangerous ones.
You hire. Then the app shows you what you actually got. The score is the accuracy of your read.
A voice conversation dies at about four seconds of dead air, so the whole design attacks latency: hold-to-talk removes end-of-speech detection, the reply streams and is spoken sentence by sentence, and the model is prewarmed while you are still talking.
Speech recognition, reasoning and speech all run on device. Nothing a candidate says or hears is transmitted anywhere.
Milestone 0 is the latency spike. Nothing else gets built until the measured number says the concept survives.
Privacy
The full policy is on the privacy page. The short version:
Pricing
Questions
No date. The first milestone is a measured latency spike on real hardware — if a conversation cannot feel like a conversation, the idea does not survive.
No. Speech is transcribed on the device and never written to a file or uploaded.
