Subprocessors
Last updated:
Regional AI in new applications
New applications with an explicitly selected AI region use Google Cloud Speech-to-Text (Chirp 3) for voice audio and GPT-4o mini through Vercel AI Gateway for transcripts and necessary field descriptions. The application sends speech to the selected EU or US endpoint and pins model inference to that same region. The Gateway itself can process requests outside this region. Earlier applications retain their previous providers until separately migrated. Gateway requests retain ZDR and disallow prompt training. These Gateway settings do not apply to direct Google Cloud requests. Regional speech uses inline requests without uploading audio to object storage; Google Speech data logging must remain disabled in the configured Cloud project. Region selection does not determine the location of all infrastructure, operational metadata, backups or support. The Article 9 and Article 10 GDPR prohibitions in the Terms and DPA applies to all these processing routes, including direct Google Speech and Vercel AI Gateway. Provider security, ZDR and region settings do not lift it.
Google Cloud — regional speech recognition
For applications with an explicit EU or US AI region, Google Cloud Speech-to-Text (Chirp 3) receives voice audio at the corresponding regional endpoint. Audio is sent inline without an object-storage upload. Gateway ZDR does not apply to this direct integration. Google Speech data logging must remain disabled in Belgi's configured Cloud project. The selected speech endpoint does not determine all operational metadata, support or backup locations.
Scope
These providers support Belgi. Their role depends on the processing: they are subprocessors where they process customer-controlled data on Belgi’s behalf. Authentication providers also process information under their own policies. Last updated: November 5, 2026.
Vercel — hosting
Hosts Belgi and published applications. Published application server regions are selected at publication. Processes deployed form definitions, configuration, requests and technical metadata. Gateway processing is described separately. This does not guarantee that all logs, backups, support or provider operations remain in the selected region.
Neon — PostgreSQL databases
The creator database is hosted in Frankfurt, Germany. It holds creator accounts, drafts and deployment administration data. Separate published application databases store users, invitations, sessions, activation information and submissions in the region selected at publication.
Vercel AI Gateway — source-document analysis
Source documents are analysed by GPT-5.6 Sol through Vercel AI Gateway using inline PDF or image input, with response storage disabled (store: false). Every document-analysis request requires Zero Data Retention (ZDR) and disallows prompt training. If no eligible provider is available, analysis fails without a direct OpenAI or non-ZDR fallback. Gateway selects an eligible inference provider; the model identifier does not alone identify the recipient. Document analysis is configured to require model inference in the EU, independently of the published application’s AI region. Belgi checks the serving region reported by Gateway before accepting the result. If an EU route satisfying ZDR is unavailable or the reported region cannot be confirmed, analysis fails without switching to another region. This restriction concerns model inference, not the entire network path, Gateway operations or provider safety-review exceptions. ZDR does not remove Belgi’s stored drafts, operational metadata, applicable legal exceptions or international-transfer obligations.
Vercel AI Gateway — voice routing
Voice input is optional; manual entry is available. Recording requires microphone permission and a compatible browser. In applications configured with an EU or US AI region, audio is sent to the corresponding Google Cloud Speech-to-Text endpoint (Chirp 3); transcripts and necessary field descriptions are sent to GPT-4o mini through Vercel AI Gateway with inference pinned to that AI region. In legacy applications without an AI region, Gateway routes audio to MAI-Transcribe 2 and transcripts with field descriptions to GPT-OSS 20B inference providers. Review and correct extracted values before submission. Browser microphone permission does not replace the application owner's privacy notice, lawful basis or any consent required to record another person. The current document-analysis integration and updated Gateway voice integration require ZDR and disallow prompt training on every Gateway request. Gateway routes only to eligible providers, including fallback routes; if no eligible provider is available, processing fails without a direct or non-ZDR fallback. These requirements apply independently of any optional team-wide Gateway setting. They do not apply to direct Google Cloud speech requests, stored drafts or submissions, encrypted retry results, operational metadata or backups. Earlier published applications may require an update. Provider terms and applicable legal exceptions still apply; ZDR does not establish a lawful basis or an international-transfer mechanism.
Microsoft Azure — voice transcription
For legacy applications without an AI region: Receives audio through Vercel AI Gateway for microsoft/mai-transcribe-2 transcription under Gateway’s ZDR routing requirements. Processing is not pinned by Belgi to the application’s region.
GPT-OSS inference providers — field extraction
For legacy applications without an AI region: Receive transcripts and necessary field descriptions through Gateway for openai/gpt-oss-20b. DeepInfra served the verified integration request; Gateway can select other eligible ZDR providers. This is dynamic routing, not a commitment to DeepInfra alone. Contact Belgi for the current recipient and contractual arrangements for your deployment. The OpenAI model identifier does not identify the company serving each request.
Resend — operational email
Processes recipient addresses, support messages, application activation links, password-reset links, password-change notifications and delivery metadata. Administrator-created invitations are shared by administrators rather than automatically emailed by that feature. Processing may take place in the United States.
Upstash — rate limiting infrastructure
Redis infrastructure is configured in Frankfurt, Germany (eu-central-1). Belgi uses rate limiting for support, sign-in, document analysis, publication and draft or application changes. It processes IP-derived keys for public requests and creator account identifiers for authenticated operations, with request counters and time windows. Form content, documents, audio, passwords and support message bodies are not included in rate-limit keys.
Google and GitHub — external sign-in
Process authentication requests, account identifiers, email addresses and profile data when their sign-in option is used. Their infrastructure is global. They also act under their own privacy terms; they are not automatically subprocessors for every processing activity.
Transfers and updates
For transfers outside the EEA, applicable safeguards include standard contractual clauses or another valid GDPR mechanism. A hosting region does not establish the location of every backup, log or support operation. Belgi maintains this list and informs customers of intended subprocessor additions or replacements under the DPA. Contact begaiym.adylbek@gmail.com for processing arrangements.