What the AI does
Authorized typed teaching uses Groq with a bounded conversation window, current board and explicitly attached source passages. The application validates a declarative visual protocol and checks suitable numeric/choice answers against a reference. These checks do not guarantee the model’s reference is academically correct. Optional narration uses ElevenLabs TTS; real-time voice uses ElevenAgents and its distinct configured hosted model.
The public learning demonstrations are scripted, deterministic examples. They do not start paid AI sessions, listen through a microphone or use a real student’s records.
Accuracy has a boundary
Earlier checked linear-equation lessons are preserved. Universal academic AI conversations support broader exploration, not guaranteed facts, complete subject coverage or reviewed curriculum alignment. Timelines, process relationships, charts, code and mathematical visuals have bounded renderers; not every scientific diagram or map is implemented. The AI can misunderstand, hallucinate or assess an open response incorrectly.
You can ask for a simpler explanation, revisit a foundation, check both sides or compare with a teacher. A successful answer does not prove mastery. Feedback and optional concept checks are limited educational evidence, not a diagnosis or a guaranteed learning outcome.
Safeguards that are implemented
Recognisable harmful/private/prompt-injection and emotional-dependency text is screened before ordinary text teaching. Structured responses are bounded and checked; model-generated code is not executed. Board tools accept only supported inputs and current lesson/session context.
The application enforces identity, ownership, request/usage limits and reviewed feature availability. These controls reduce specific risks; they are not a promise of comprehensive moderation, flawless accuracy or perfect security. Novel wording, languages and voice interactions need further evaluation.
Student voice is not broadly released
A published provider agent and passing test examples are not written permission for use by minors. The current inspected agent’s streaming guardrails and transcript-deletion settings still need review. Student voice admission is restricted until the required safeguards, access states and provider/legal permissions are verified.
Six synthetic provider text safety evaluations passed in the prior technical audit. That is not a successful live student conversation, full safety validation or proof of microphone shutdown on every device. The private owner-test workflow remains separate from student release.
Your microphone is a choice
No microphone starts because you opened or scrolled this website. Where live voice is authorised, a deliberate connect action requests browser permission. You can end a call and return to ordinary learning. The application requests shutdown on leaving or hiding the classroom, but production/device acceptance still has to verify a real call.
Voice processing involves ElevenLabs and its hosted conversational model. Audio recording and transcript storage are different settings. Read Privacy before choosing an available sensitive feature, and leave personal details out of a question.
Keep human relationships human
The teacher is not a human friend, romantic partner, therapist or guardian. It should not ask for secrets, exclusive trust, private contact or emotional dependence. It should not shame a student or pressure them to continue.
Personal distress and immediate danger need real-world help. Tell a trusted adult, teacher or school counsellor; contact appropriate local emergency help in immediate danger. Zuvopa is not a crisis service or a substitute for medical, legal or other professional support.
Report a problem without sharing more than needed
You can report incorrect mathematics, harmful or inappropriate responses, board/audio problems, privacy concerns and accessibility barriers. Signed-in lessons have category-based private reporting. The public Contact page offers corresponding routes when its secure queue is operational.
A receipt means stored for review, not a live human response or resolution. Reports are not public. Do not attach private recordings, passwords, identity documents or another student’s information. The operator must staff review and escalation before launch.
What is still being checked
The actual production database, provider key restrictions, legal permissions, provider retention, supported-device audio behaviour and end-to-end deletion remain release-review items. Public documentation does not turn an unverified feature into an authorised one.
Zuvopa should invite curiosity without pretending to know everything. If a response is wrong or uncomfortable, stop and ask a trusted person. You do not owe the AI a conversation.
A question or a concern?
Ask about this information, request support, or report something that needs attention.
Contact Zuvopa →