TherapyNow
About this project
In one sentence, what are you building?
Outbound AI call centers for Masky avatars: give the agent a goal, it finds matching people with Exa, phone-calls them in the avatar's own cloned voice over Vapi/Twilio, walks them through a human-authored step-by-step script — and after every single thing the callee says, a calibrated TypeSafe probability judge (Jev, not an LLM) re-scores the odds the call will still reach its goal and gracefully ends it the moment that drops below 5%.
What problem are you solving?
LLM voice agents are unaccountable on outbound calls: they improvise, hallucinate, and don't know when to stop — disqualifying for high-stakes outreach. We flip the architecture: the human author owns every word (a deterministic script engine — the agent can only say what you wrote) and AI is demoted to judging, not speaking: keyword judgments advance the script, and a numeric P(goal) verdict after each utterance decides continue vs. hang up. Our live demo is the hardest version of the problem: proactive crisis outreach — find a person showing signs of depression online, call them in a trusted therapist-twin's voice, stay on the line until they're signed up on therapynow.masky.ai and matched with a REAL human therapist — and if the probability judge says the call is doing more harm than good, it exits gracefully with the 988 line instead of pushing on. Built tonight on our existing inbound call-center flow: same scripts, same voices, new direction.