A subprocessor is a company that processes personal data for Berri. Storage of chats and AI replies is in Mumbai. Some of the companies below process data outside India when you use that feature.
- Amazon Web Services
- Hosting, database, accounts, AI replies on Amazon Bedrock, transcription, short lived files, logs, reminder schedules, and email sending from Mumbai (ap-south-1). The public website may be cached on CloudFront, which is a global network. Chat text is not the page cache.
- AI models
- Replies come from open weight models that Amazon Bedrock hosts and runs in Mumbai (ap-south-1). The defaults in the code are Qwen3 235B and Qwen3 Next 80B (made by the Qwen team at Alibaba Cloud), DeepSeek V3.2 (made by DeepSeek), and Mistral Large 3 (made by Mistral AI). Berri sends each request to Amazon Bedrock only and has no direct connection to those model makers. Model selection can vary with configuration.
- Meta
- WhatsApp through the official Cloud API, and Instagram or Messenger if those channels are on. Meta processes the message and the identifier as the channel. Berri does not use unofficial WhatsApp tools and does not send cold messages.
- Optional. Off until you connect. Default read only mail and calendar. Also, a public Google Sheet you ask Berri to read, and sign in if the owner enables the Google button. Not used to train AI.
- Microsoft
- Optional. Off until you connect. Default read only Outlook mail and calendar. A send or a calendar change needs a separate explicit yes. Not used to train AI.
- Telegram and Slack
- Only if those channels are switched on and you message Berri there. They process the channel traffic.
- Search
- When a question needs the web, the query (not your saved profile) may go to Brave, Tavily, DuckDuckGo or Wikipedia.
Berri does not sell personal data and does not train AI on your chats. Questions go to the grievance page.