Regions below are what the provider pages said on 1 Nov 2026, plus the region this project is configured to use. They aren't a claim that every vendor is inside the EU, and they aren't a claim that no vendor trains on data.
Supabase Pte. Ltd.
Role Database, sign-in storage, and the API. Also Ownflow's model storage, the second source of the speech models, used without an account too.
Region This project is in Frankfurt (eu-central-1). The Supabase DPA (version dated 1 June 2026) says data is stored and primarily processed in a region the customer directs, and that Supabase may still process it where a subprocessor or the law requires. The subprocessor list of 12 Oct 2026 names US hosting companies, including Amazon Web Services, Google and Cloudflare.
Training and retention The terms fetched 1 Nov 2026 say Supabase won't use customer data to train a model without prior written consent. The free plan has no daily hosted backup. Nightly dumps run on the operator's Mac. See the retention table.
Deepgram
Role Turns meeting audio into text. Ownflow calls api.eu.deepgram.com with mip_opt_out=true.
Region Deepgram's current data page says that EU host stays in the EU and doesn't fall back to another region. The specific EU country can change. Deepgram's privacy policy, last updated 26 Oct 2021, still says it stores data on servers in the US. Both sentences are on Deepgram's site. A short-lived token doesn't itself force the EU host or the opt-out; the app sets both.
Training and retention Deepgram's default is to keep audio and transcripts for model improvement. The opt-out flag is how Ownflow turns that off per request. Opted-out audio is described as kept only while the request is processed. Usage logs are still kept, and Deepgram says those logs are retrievable for 90 days. Deepgram's subprocessor page lists audio processors including Amazon Web Services, Oracle, Cloudflare, OpenAI and Anthropic.
Groq
Role Cleanup, rewrite, meeting notes, and screen questions. A screen question sends the screenshot and the question. The vision model is qwen/qwen3.8-27b. Cleanup, rewrite and notes tests use openai/gpt-oss-20b. That name isn't evidence that OpenAI sees the text.
Region Groq's data page says retained customer data is in Google Cloud buckets in the United States. The DPA allows processing in the US and other countries, with standard contractual clauses as the stated transfer tool.
Training and retention The services agreement, last modified 22 June 2026, says Groq may not train or fine-tune on inputs or outputs unless the customer allows it. When zero data retention is on, it stops Groq from logging inputs and outputs for reliability and abuse checks.
Google
Role Sign-in only, for this beta. Scopes are openid, email and profile. The Google app is in Testing, for at most 100 test users, on GCP project ownflow-513418.
Region Not pinned to Frankfurt by Ownflow.
Training and retention Not restated here. Google isn't the model provider.
GitHub
Role Hosts the app downloads, the update feed and the first copy of the speech models, in Ownflow's public ownflow-releases repository. Also used without an account.
Region Not pinned to Frankfurt by Ownflow.
Training and retention GitHub receives the download requests, with your IP address. It gets no dictation, account or meeting data from Ownflow.
The evidence, with the page addresses, is in the processor register that ships next to this file in the Ownflow source. Provider lists change. The provider pages were read on 1 Nov 2026. The 2 Nov changes are the vision model, zero data retention, the Mac backups, and account export and deletion.