Understudy

Voice-AI rehearsal and assessment platform for the conversations that matter. The AI plays any counterpart (a customer, a patient, a passenger, a candidate or a colleague) and reacts to how the participant behaves, tuned by difficulty and patience dials; then an LLM judge scores every rubric criterion with a verbatim quote that is checked against the transcript. A three-act studio (author, rehearse, review), a manager dashboard, an admin console and a participant join flow with cohort invite codes: participants enter by name and never see their scores, while managers get full scorecards and CSV export. A scenario compiler turns a short brief into a persona and a rubric; the LLM (DeepSeek by default, Claude optional), the data store (Supabase with Row Level Security) and the voice (Deepgram speech to text, ElevenLabs text to speech) are all swappable. SCORM 1.2 export, HubSpot CRM push, an MCP server with six tools and agentic onboarding through POST /onboard. A monorepo with a Next.js 15 and React 19 front end and a FastAPI back end (39 endpoints); about 440 automated tests run in CI on every push; English and Turkish interface with versioned KVKK consent.