AI grades the lesson.
Mathematics proves the score.
One recorded lesson goes in. A complete, defensible evaluation comes out, and every step can be checked.
Three layers.
One defensible score.
Each layer does one job, and no layer can quietly override another.
Intelligence reasons
The AI reads the lesson across languages and scores each strand against cited evidence. It never touches the math, and it never hears the raw audio.
Computation verifies
A statistical audit sits between reasoning and publication. It runs on code, not language models, and measures how sound each score is before anyone sees it.
Educators set the standard
Experts calibrate the rubric. Anything the audit doubts goes to a human reviewer. Doubt goes to a person, never silently into a file.
Layer 1 reasons. Layer 2 audits. Layer 3 contextualises.
The AI judges.
Mathematics checks its work.
An AI can sound just as confident when it's inventing as when it's right, so we never take its word for it.
Consistency
Every judgement is read many times over; a score that wobbles is held, never published.
Grounding
Every claim must trace to a real, locatable moment in the audio, checked by signal processing that cannot itself hallucinate.
Baseline
Each score is weighed against the teacher's own pattern, so anything unusual surfaces for a human.
We're open about what's live today and what's still being built, and we say which is which. Honesty is part of the product.
Passes the audit? Published.
In doubt? A human sees it first.
Published
A lesson that clears every applicable check becomes a verified report, evidence and all.
Quality review
Anything the audit is unsure about is routed to review before it goes near a teacher's file.
Human review
A score the audit doubts is held for a human educator. The decision stays with a person.
A complete report,
not a lonely number.
Behind every score is a rubric forged over thousands of real classroom observations, then made to prove itself, moment by moment.
Timestamped quotes
Every finding cites the exact moment it came from, in [00:17] form, with the excerpt.
Language analysis
Xyova tracks switching between Urdu, English and Arabic, including frequency and tone, mid-sentence.
Speaker analysis
Speaking-time split, students anonymised as "Student 1 / Student 2". Zero names, ever.
Longitudinal profile
Sessions over time, a personal baseline, moving average and "Personal Best" labels.
Verified, in every report
Each report carries proof it passed an independent layer of statistical checks.
A human gold standard
Experts calibrate the standard and review anything in doubt. It builds a benchmark that can't be bought, only earned.
Built for classrooms that switch
language mid-sentence.
English, Urdu and Arabic in a single breath. Single language tools mishear these lessons. Xyova was multilingual from the first line of code.
Bring one recorded lesson.
See the proof.
Request a demo and evaluate a lesson of your own.