Loma AI beta: service providers (subprocessors)
Version 2026-11-20.2 · Last checked 2026-11-20 (Anthropic only)
These companies process personal data for the Loma AI beta on our behalf. We tell students before adding one that would get lesson content.
| Provider | What it does | Data it gets | Where | Retention at the provider | Data use / training | Terms checked (2026-11-16) |
|---|---|---|---|---|---|---|
| Anthropic, PBC (Claude API) | AI model that writes Loma's replies; also used to help analyze lesson records, through Claude Code on our Claude Team plan (same Commercial Terms; Anthropic keeps Claude Code data 30 days, and Claude Code keeps its session history on our computer, which we delete within 30 days) | Conversation transcript, screen state, answers, lesson context; pseudonym only, no name or email. Homework files students upload (images or PDFs), sent once inline in a Messages API request (never the Files API) to copy out the problems; a file may show what the student didn't crop out | Data stored in the US. By default requests may be routed to the US, Europe, Asia or Australia unless a region is set | Deleted within 30 days; if flagged by Anthropic's safety systems, up to 2 years (classification scores up to 7 years) | Commercial terms § B: "Anthropic may not train models on Customer Content from Services." DPA incorporated (§ C) | Commercial terms · DPA · Training · Retention · Location · Usage policy · Claude Code data use |
| LiveKit (LiveKit Cloud, LiveKit Inference) | Hosts and runs Loma's voice agent; carries live voice; routes speech-to-text and text-to-speech. No lesson audio is recorded (Egress is not used since 2026-11-16.2) | Everything in a lesson as it happens: both sides' audio, transcripts, screen state, the AI's instructions, under the pseudonym; agent technical logs (no transcript text) | US: agent deployed in us-east; LiveKit processes personal data in the US (privacy policy) | Media "deleted after delivery" unless recording or logging is enabled. LiveKit's session recording is off for Loma AI (record: false). Inference: zero data retention by default. Agent runtime logs (no lesson content): kept for a limited time; LiveKit's logs docs don't state a runtime-log period and point to the session-data window (30 days), which we don't rely on since session recording is off (checked 2026-11-16). Build logs (no student data) are kept indefinitely | ToS § 7.1: will not use Customer Data for model training or fine-tuning | ToS · DPA · Privacy · Sub-processors · Inference · Observability |
| Deepgram (via LiveKit Inference; LiveKit's sub-processor) | Speech-to-text (Flux), in lessons and for spoken end-of-lesson comments | Student's voice audio | US (LiveKit sub-processor list) | Zero retention per LiveKit's contract with it | LiveKit says it contractually requires inference sub-processors not to retain, log or train on Inference Data. Deepgram's own public default is to train unless opted out (mip_opt_out); LiveKit's contract, not that default, governs this traffic | LiveKit sub-processors · Deepgram privacy · Deepgram MIP |
| Cartesia (via LiveKit Inference; LiveKit's sub-processor) | Text-to-speech (Sonic) | Loma's reply text (no student audio) | US (LiveKit sub-processor list) | Zero retention per LiveKit's contract with it | As for Deepgram. Cartesia's own public terms default to training unless opted out; LiveKit's contract governs this traffic | LiveKit sub-processors · Cartesia privacy · Cartesia terms |
| Fly.io | Hosts the Loma AI session server (lesson state, math checking, recording intake, sign-up and Settings API) | Everything the server handles; server logs with IP addresses and pseudonyms (no transcript text) | US: region iad (Ashburn, Virginia) | App data not kept by Fly beyond our machine and logs. Logs: "We currently retain logs for 7 days" (docs, checked 2026-11-16). Volumes encrypted at rest | Processor on our instructions; no AI-training statement found | Privacy · Security |
| Supabase | Postgres database (accounts, consents, lesson records, deletion log). Storage held opt-in lesson audio until 2026-11-16.2; no new audio is stored and the server's daily sweep deletes what was kept | All stored beta data | AWS us-east-1 (North Virginia) | Until we delete it. Database backups: up to 7 days ("Pro Plan projects can access the last 7 days of daily backups"; docs, checked 2026-11-16; plan to be confirmed by the founder). "Database backups do not include objects you store via the Storage API" | DPA § 3.3: no use for purposes other than the agreement | Privacy · DPA · Security · Backups |
| Vercel | Hosts loma.study (sign-up and legal pages) and the Loma AI web app; Vercel Web Analytics | Website requests and IP addresses; analytics: page URL, referrer, country/region/city, OS, browser, device type, with a request-based visitor hash and no cookies; custom event counts (the step, a yes/no or right/wrong, and whether the visitor was signed in, never who) | Primary processing in the US (DPA) | Analytics visitor sessions discarded after 24 hours; platform backups 30 days | Hobby/Pro plans: Vercel may share de-identified data with AI partners unless "Opt-out of Model Training" is set in team settings. Set it for both projects before publishing | Privacy · Analytics privacy · DPA · Security |
| Clerk | Accounts and sign-in for loma.study and Loma AI | Name, email, account ID, sign-in sessions and cookies, device/browser and IP data for sign-in security | Primarily Google Cloud and Cloudflare; may be stored "anywhere" Clerk or its sub-processors operate | Deleted within 90 days of contract end (DPA § 9); account data while the account exists | Processes customer personal data only on our instructions (DPA § 4.2); no AI-training statement found | Privacy · DPA |
| Resend | Sends beta emails (access, notices, deletion confirmations) | Email address, email content, delivery logs | Stored in the US | Email and log data kept 30 days | "Neither Resend nor these subprocessors use customer content to train or fine-tune models." | Privacy · DPA · Security · Sub-processors |
| Sentry | Error reports from the loma.study website's pages, including the Loma AI sign-up and legal pages at loma.study/ai (loma.study already used it for the diagnostics). The Loma AI app, server and voice agent don't use Sentry, so it receives nothing from lessons | Technical error details (scrubbed: no names, emails or lesson content; no session replay) | US (*.ingest.us.sentry.io) | Error events: 30 days (Developer plan) or 90 days (Team/Business), per Sentry's retention docs (checked 2026-11-16); plan to be confirmed by the founder | Not checked this round | Privacy (not read) |
Text-to-speech provider
Production is pinned to Cartesia. LomaAR has a config option for Inworld (also a US LiveKit inference sub-processor). Inworld's own terms were not checked. Before switching: check them, add a row, publish a new version of this list, and, because the sign-up's terms checkbox has students agree to processing by the providers named in the privacy notice, update the privacy notice, publish a new consent text version and ask students to agree again.
Not used in the beta
Langfuse, Mathpix, Postmark and Better Auth appear in LomaAR's long-term infrastructure plan (docs/INFRASTRUCTURE.md) but are not part of this beta. GitHub hosts Loma's code; by policy no student's words go into the code or tests, so it receives no lesson content and no student's name, email or code. Review notes kept with the code contain only totals and scrambled lesson numbers that can be matched to a lesson only with our own records. The browser's own speech recognition is not used (spoken comments go through Deepgram via LiveKit). If any of this changes, update this list first.
How the checks were done
Pages were read on 2026-11-16 by research agents and an independent auditor, and quoted. Not verified: Clerk's SOC 2 status and security page, and whether Clerk's PostHog session replay covers Clerk components on loma.study; whether Vercel's model-training opt-out is set; the mechanics of Vercel's visitor hash; whether LiveKit sends Deepgram's mip_opt_out flag; Inworld's terms; Cartesia's data-controls page (rate-limited); Sentry's terms; Fly.io's DPA text. Supabase privacy and Vercel DPA were partly summarized rather than read in full. See REVIEW-NOTES.md §5.
