The practice matches the real thing — the record matches what happened.

A large part of the question bank is templated: a question type with a formula and value ranges, instantiated from a seed. Two students choosing the same category get different papers; retaking gives you a different one too. Non-templated questions mix in alongside.

You pick a category, a difficulty and a length. The server assembles the set, fixes a deadline, and grades server-side only. The result is a score, a per-topic breakdown, and a review with the correct answer, an explanation, and your time for every question. If the deadline passes, it auto-submits whatever you had.

Pick a role or paste a job description and your experience level. The AI builds an interview plan, then speaks each question aloud with the text on screen. You answer by voice — the browser transcribes it — or type where voice input isn't supported.

After an answer, the AI decides whether a follow-up would help and, if so, asks one that references what you actually said — not a fixed next question. Your camera records locally so you can watch your delivery back before the report, which scores communication, structure, technical depth and confidence, with per-answer notes and a model answer. With AI off, a fixed question bank runs the same flow with rubric feedback.

Proctoring applies to mock tests and AI interviews, never to a live training class. Before a proctored attempt you see a pre-check: camera preview, a short checklist, and a plain-language list of exactly what is recorded — then you consent.

During the attempt, in the browser only: tab and window switches, leaving fullscreen, pasting into an answer, the camera being stopped, and — where the browser has a face detector — no-face and multiple-face events. It does not track gaze or analyse behaviour, and it never blocks you. Every event goes into an integrity report you can see for your own attempts.

A problem statement, constraints, examples and starter code in JavaScript or Python. Run against the visible test cases in a Web Worker with a hard timeout; submit to run the hidden cases too. It's clear about the limits — practice, not a sandboxed judge — but it's the shape of a real online-assessment round, and it's proctored when launched from an assessment.

A trainer opens a session for a batch; students join a real-time room from their own devices. Camera is opt-in every session, and that consent is enforced where the media token is issued — not just hidden in the UI. Large classes default to listen-only, and the trainer can hand one student the floor.

During the lecture a multiple-choice question is pushed to every screen at once, under a countdown whose end time the server sets. Each student submits one answer — a database constraint, not an application check. The correct option and the class breakdown are withheld until the timer closes and never leave the server before that, for any role. When the session ends, attendance — weighted by minutes present and questions answered — and the session record compute themselves.

For a student: progress, accuracy, topic strength, points and badges. For a trainer: per-session and per-batch reports, plus the proctor reports for their students' attempts. For the placement cell: org-wide KPIs with previous-period deltas, batch performance, weakest topics first, and an at-risk list where every flag carries its reason.