Subprocessors & models
Every third party that touches data
This list is maintained from the codebase's configuration — if a service isn't configured in the app, it isn't on the list, and vice versa. Last updated October 20, 2026.
| Vendor | Purpose | Data touched | Region |
|---|---|---|---|
| Vercel | Application hosting (web app + API routes) | All application traffic in transit: educator accounts, student names/emails, transcripts, scores | United States (single-region deploy) |
| Supabase (PostgreSQL) | Primary database | Educator accounts, quizzes, rubrics, student names/emails, transcripts, AI + professor scores, evidence excerpts | Not pinned in code — available on request |
| OpenAI (API) | Speech-to-text (gpt-4o-mini-transcribe); the AI examiner (gpt-4o, with gpt-5.4 for live-probing follow-ups); rubric scoring (gpt-5); educator tools and essay-defense question generation (gpt-4o, gpt-5-mini). Most of these are defaults that configuration can override. | Student speech audio, transcripts, the student's essay on essay-defense assessments, question + rubric text, instructor personas/notes, instructor-uploaded course documents. Never webcam/screen video. | Not specified by vendor for standard API traffic |
| LiveKit Cloud | Realtime voice/video rooms for the assessment session; recording egress; hosts our examiner agent; routes the examiner's speech synthesis to Inworld | Live microphone audio (plus webcam/screen on video-enabled quizzes), room names, session metadata | Distributed cloud (edge media routing) — not pinned in code |
| Inworld (via LiveKit) | Text-to-speech for the examiner's voice, reached through LiveKit's inference gateway — we hold no direct contract with Inworld | The examiner's spoken lines only — the question text, follow-ups and closing it reads aloud. In essay defence those lines quote passages of the student's own essay. Never student speech, never recordings. | Not specified by vendor |
| Amazon Web Services — S3 | Recording storage (private bucket, short-lived signed URLs) | Audio recordings, webcam/screen video files, question audio prompts | United States (us-east-1) |
| Amazon Web Services — SES | Transactional email: account verification and password-reset links, retake and alternate-email verification codes, submission notifications to educators, and demo-request notifications to the founder | Recipient email addresses and what each message carries: sign-in links and one-time codes for educators and students; the assessment title and the submitting student's name (first-submission notice) or submission counts (daily recap) for educators; demo-request contact details for the founder. Bounce and complaint notices about those addresses come back through Amazon SNS. | United States (us-east-1) |
| Google (sign-in) | Optional sign-in (“Continue with Google”) for educators and students | Only for people who choose “Continue with Google”: Google authenticates them, returns their name, email address and basic profile to us, and learns that they signed in to QPAssess. No assessment content, recordings, transcripts or scores are sent to Google. | Google's global sign-in service — not pinned in code |
| Microsoft (Entra ID sign-in) | Optional sign-in (“Continue with Microsoft”, including school Microsoft 365 accounts) for educators and students | Only for people who choose “Continue with Microsoft”: Microsoft authenticates them, returns their name, email address and basic profile to us, and learns that they signed in to QPAssess. No assessment content, recordings, transcripts or scores are sent to Microsoft. | Microsoft's global sign-in service (Entra ID) — not pinned in code |
Honest footnotes
- Trigger.dev appears in our dependencies but is not active: no project is configured and no data flows to it today (scoring runs inside our own infrastructure). If we activate it for background scoring, its payloads are limited to internal record IDs — never transcripts or names — and this page will be updated first.
- Canvas (or your LMS) is not a subprocessor — it's your institution's own system. QPAssess sends professor-reviewed grades back to it only when the professor triggers grade sync via LTI.
- Model-level detail on what each OpenAI model sees and decides is on How AI is used.