The research platform for empathic AI
Every experiment we publish runs on infrastructure we built ourselves. Participants talk to a live, working AI agent, the emotion in their voice is measured as they speak, and a threshold you set can change how the agent answers. Built for our own studies, and open to other research groups.
What happens in one call
- 01
The call runs on our own pipeline
The participant speaks to a working voice agent. The platform runs the listening, the replying and the speaking as one pipeline of its own, so which service does each part is a choice you make rather than something you inherit.
- 02
Every turn is scored
The emotion in the participant's voice is measured as they speak, turn by turn, and the agent's reply waits for that score rather than arriving beside it.
- 03
A threshold can change the reply
You set what to watch and how far it has to move. When a turn crosses it, the agent either says a line you wrote or answers with empathy for the next two replies, and the change reaches the voice before it is spoken.
- 04
The manipulation is data
Each firing is written to the conversation with the rule that caused it, so what the agent did and why is part of the record rather than something to reconstruct afterwards.
What you can run on it
Voice conversations
Participants talk out loud with a live, empathic voice agent, in the same interface whichever AI service runs it.
AI interviewer
A voice interviewer works through your questions, and a participant whose call drops can rejoin where they left off.
Text chat
Participants type instead of talking, and every message is kept for analysis.
Multi-session studies
Participants come back for a series of sessions, and the agent can remember what was said before.
Built so the method survives the software
A study runs for weeks, and the services underneath it do not hold still for weeks. The platform is built so that nothing about a live condition changes unless a researcher decides it.
- 01Six AI providers behind one identical participant interface, so the service is never a confound.
- 02Voice prosody scores persisted with every message, as research data rather than telemetry.
- 03Emotion-triggered experimental manipulation, with every trigger firing logged as data.
- 04Recruitment-aware participant IDs, concurrency caps with a fair waiting queue, and consent gates.
- 05Daily drift monitoring that reports provider changes instead of silently substituting models mid-study.
Run your study on it
The platform is open to other research groups. Tell us what you want to run and we will tell you whether it fits.