Trust center
How QPAssess handles your students' data
QPAssess runs AI-moderated oral assessments: a student speaks with an AI examiner, the conversation is transcribed and provisionally scored against the instructor's rubric, and the instructor reviews everything before any grade counts. This page describes — plainly and without certification badges we don't have — what data the product collects, where it lives, who can see it, and what we have not built yet. Last updated November 20, 2026.
What we collect, and from whom
From students
- Name and email, typed at the start of a quiz (or checked against an instructor-provided roster when the instructor restricts access). Taking a quiz needs no account. Students who choose to create one also have an account email, a password stored hashed if they set one, and any alternate email addresses they add.
- Microphone audio of the assessment conversation, and the resulting transcript.
- Webcam and/or screen video, only if the instructor turns that mode on for a quiz — and only after the student sees a consent notice and clicks through it.
- Scores and feedback: the AI's proposed per-criterion scores, supporting transcript excerpts, and the instructor's final scores and notes.
- The network address the attempt connected from, recorded once when the attempt starts, so an instructor can see whether several attempts arrived from one place. Only the instructor who owns the assessment can see it; it is never shown to other students, and nothing automatic is decided from it.
From educators
- Name, email, institution, an optional role (e.g. professor or teaching assistant), and a password (stored hashed by Better Auth, never in plaintext). An educator who signs in with Google or Microsoft needs no password; for that sign-in we keep the provider's account identifier and the sign-in tokens it returns.
- Login session records, including IP address and browser user-agent, for session security.
- For Canvas-connected accounts: the Canvas instance URL and which courses are synced. We do not store per-educator Canvas API tokens.
- Activity log: changes to courses, assessments, course templates, memberships, branding and grades are recorded with who made them and when — references and titles, never question text, transcripts or essays. Entries are deleted after 13 months. An institution's admins see only what came from the institution, and never which student a grade change concerned; the professor who owns the course sees that. A person's name stays on the entries they made until those entries are deleted.
- Canvas launch log: each time someone opens QPAssess from Canvas, we record, when available, which Canvas installation it came from, the page it opened, whether it was accepted and, if not, a short reason. We never store the sign-in token. We use it to diagnose Canvas setup problems. Entries are deleted after 13 months.
From website visitors
The landing page collects only what you type into the demo-request form (name, email, institution, role, optional message). There is no third-party analytics or ad tracking on the product.
Where data lives
- Structured data (accounts, quizzes, rubrics, transcripts, scores) lives in a PostgreSQL database hosted by Supabase.
- Recordings (audio, and webcam/screen video where enabled) live in a private Amazon S3 bucket. Playback links are signed URLs that expire after 15 minutes; the bucket does not allow public reads.
- A full list of third parties that touch data — with what each one sees — is on the subprocessor & model registry page.
Encryption
- In transit: browser↔app traffic is HTTPS; realtime audio/video to LiveKit uses encrypted WebSocket/WebRTC transport; calls to OpenAI and AWS use HTTPS.
- At rest: S3 applies its default server-side encryption (AES-256, AWS-managed keys) to stored recordings, and the Postgres database relies on Supabase's vendor-managed encryption at rest. We do not currently add our own application-layer encryption on top of these vendor defaults.
Authentication model
- Educators sign in with email and password via Better Auth, or with a Google or Microsoft account. Sessions are httpOnly cookies (never localStorage), marked Secure over HTTPS, expire after 7 days, and sensitive operations require a session fresher than 5 minutes.
- Email verification is enforced. A new email-and-password account — educator or student — cannot sign in until its owner confirms the address from a link we email. Accounts created before verification was switched on (September 2026) were marked verified rather than locked out.
- Students don't need an account to take a quiz. When a quiz session starts, the server issues a short-lived signed token bound to that one submission; later requests (and the realtime room join) are verified against it, so a quiz link alone can't be used to read or write someone else's work. Students who create an account — for the student portal or the practice lab — sign in the same way educators do.
- Canvas / LMS launches use LTI 1.3: signed JWTs verified against the platform's published keys, an issuer allowlist that fails closed, single-use nonces, and a deployment allowlist.
Recording lifecycle
- Capture: audio is recorded during the assessment conversation. Webcam and screen capture happen only on quizzes where the instructor enabled them, after the student passes an explicit consent screen.
- Storage: recordings are written to a private S3 bucket (one audio file per question; separate webcam and screen files when video mode is on).
- Access: only the instructor who owns the quiz can play recordings, through their authenticated review dashboard. Playback URLs expire after 15 minutes. Students cannot replay their own or others' recordings.
- Deletion — honest version: a professor can delete a student's attempt from the review screen, which removes its transcript and scores and the app's ability to play its recordings. The recording files themselves stay in the private bucket until we delete them by hand on request. Automatic deletion of recordings and transcripts 12 months after an assessment closes is built and being switched on in stages; until it is on, there is no automatic retention schedule. See the privacy policy.
Optional ID check
- A professor can turn on an ID check for an assessment. The student then holds a photo ID next to their face for one photo before the first question; if the assessment does not record the webcam, or the webcam cannot be recorded, one photo of their face is also taken at a random moment in each question. The photo is optional: students choose whether to agree before starting, and anyone who declines or skips still takes the assessment, giving a reason when they skip; the professor sees that they did.
- Nothing is computed from these photos. No face detection, no face matching, no facial recognition, and they are never sent to an AI model. The professor compares them by eye. They are stored privately in our storage provider (see subprocessors) and nowhere else.
- Only the professor who owns the assessment can see the photos, through links that expire after 15 minutes. They are not in the class list or any export.
- Retention: photos are deleted as soon as the professor marks them as matching the ID. If the professor marks them as not matching, they are kept for up to 365 days while it is resolved. Photos nobody reviews are deleted 120 days after they were taken. Deleting the attempt deletes its photos at the next nightly cleanup.
How AI is involved
OpenAI models transcribe the student's speech, hold the examiner conversation, and propose a provisional score against the instructor's rubric, with quoted evidence from the transcript. The AI never finalizes a grade: instructors review every response, can override any score, and only instructor-released grades reach the gradebook. Recordings of webcam or screen are never sent to any AI model. The full breakdown — which models, what they see, what is stored, and the vendor's data-use policy — is on How AI is used, and what the AI examiner is known to be bad at is on Known limitations.
Security practices in place today
These are engineering rules enforced in the codebase and its CI today — not aspirations:
- Every API route validates its input with a schema (lengths, ranges, ID formats) before touching the database.
- Tenant isolation: every educator request walks the ownership chain, and non-owners get a 404 (the record's existence isn't even confirmed). Automated tests assert this on every change.
- No student names or emails in URLs, and no raw-HTML injection points in the UI.
- Internal webhooks are HMAC-signed with replay protection; admin endpoints use constant-time token comparison and fail closed when unconfigured.
- Database-backed rate limiting on abuse-prone endpoints.
- Configuration is validated at boot — the app refuses to start with missing or malformed secrets rather than running half-configured.
- CI blocks merges on lint, type, and build errors and on auth-boundary test failures. Dependencies are scanned for known vulnerabilities on every change (currently warn-level, not merge-blocking, while a handful of transitive advisories await upstream fixes).
What we don't claim
- We are not SOC 2 or ISO 27001 certified, and we have not yet commissioned a third-party penetration test.
- We do not advertise "GDPR compliant" or "FERPA certified" badges. We treat assessment recordings and transcripts as part of the student's education record and say so in the student-facing consent notice; formal compliance programs come with institutional contracts.
- We run in a single region (United States). There is no EU data residency option today.
If your institution's review requires something not answered here, email jake@questionpunk.com — you will get an answer from the people who wrote the code.