CandiVision

CandiVision Roleplay turns a training lead's one-line problem ("our engineers make outage calls worse") into a voice role-play programme built on research, not guesswork. A research agent searches OpenAlex, Crossref and a vetted method library. Every skill, level and rubric line must cite a source the tools actually returned, or quote the customer's own file word for word. Anything else is rebuilt, or turned into a question for the training lead. Learners hold a button and talk to an AI character whose mood moves with what they actually do, and who reveals hidden facts only when the learner earns them. After the call, two independent Meta models judge each rubric line and quote the moment from the call. When they disagree the line shows as "unsure" for a person to settle, never averaged away. The debrief shows the moment, the source behind the rubric line, one thing to try, and a 60-second drill on the weakest must-have. It runs in English and Vietnamese, has 19 ready-built programmes, and is live at https://platform.candivision.co with no login. On launch day (2 Nov), people took 41 calls, more than half of them on phones.