Security and data protection

For IT and data protection officers – where data is processed, service providers, encryption, access and audit trail, retention and erasure, consent, recording notices and anti-discrimination guardrails.

This page describes what the product does technically, as a basis for your own assessment. It is not legal advice. The binding documents are your contract and the data processing agreement (AVV); if this page differs from them, please ask us. What the account admin can set is in Data protection.

Where data is processed

Platform. Servers, database, cache, file storage (including call recordings and applicant documents), message queues and secrets run on Google Cloud in Frankfurt (region europe-west3). The database is backed up daily, with point-in-time recovery.

Sign-in. Accounts (e-mail address, password hash, Google sign-in) are managed by Google's Firebase Authentication, a global Google service in a separate Firebase project. The code does not set a storage region for it.

Voice. Calls run through LiveKit Cloud in Frankfurt. Speech recognition uses Deepgram's EU endpoint, with use of the audio for model training switched off.

E-mail and WhatsApp agent. Nimmi's written replies are generated with Google Vertex AI on an EU endpoint. The masking described under Anti-discrimination guardrails below applies to this agent.

Processing outside the EEA. These services currently process data outside the European Economic Area or on a service that is not bound to the EU:

Service What is processed Note
Language model for phone calls the spoken conversation as text Google Gemini API, not bound to the EU; recorded as an exception. A Gemini model through LiveKit's EU endpoint answers when the primary is too slow
Summary after a call the call transcript Google Gemini API
Quality check of calls the call text an AI judge on the Google Gemini API; it stores only scores and labels
Classifying e-mails from new senders subject and up to 3.000 characters of the body of an e-mail without an OpenImmo attachment, unmasked Google Gemini API; decides whether the e-mail becomes a lead
Summary of a new lead the enquiry text, the applicant's first name and the portal's fields, unmasked Google Gemini API; Vertex AI in europe-west1 only if no Gemini API key is configured
Refine & Preview (answer to an open question) and Create Slot for This Time the applicant's first name, your answers, the listing title and address, the account owner's name and – if no viewing is booked – Nimmi's phone number for the closing lines; the preferred viewing times as the applicant wrote them OpenAI (gpt-4o-mini), when an OpenAI key is configured
Chat on the public website what visitors type into the chat Google Gemini API
WhatsApp via Nimmi's number WhatsApp messages Twilio processes WhatsApp in its US region
E-mail incoming and outgoing e-mails and addresses SendGrid (Twilio), receiving through Inbound Parse and sending through SendGrid's global API, without EU data residency
Advertising conversions (Reddit) Reddit's click ID, an amount and a currency after a purchase via a Reddit advert sent by the server without cookie consent; no name, e-mail or IP address

Voice output. Nimmi's spoken text (which can include names) is converted to speech by the configured voice. The agent's documented production setting is Gradium through LiveKit Inference; the setting itself is held outside the code repository, and the code does not state where Gradium processes the text. Google Cloud Text-to-Speech (EU location) and Cartesia through LiveKit are configurable alternatives.

Service providers

The code uses these service providers:

Provider Purpose
Google Cloud (incl. Firebase Authentication, Vertex AI, Gemini API, Text-to-Speech) hosting, database, storage, sign-in, AI models
LiveKit voice calls, voice agent hosting, access to voice models (LiveKit Inference)
Gradium (through LiveKit) voice output (documented production setting, see above)
OpenAI Refine & Preview and Create Slot for This Time, when configured
Deepgram speech recognition (EU endpoint)
Twilio phone numbers, call connection, WhatsApp sender
SendGrid (Twilio) sending and receiving e-mail
Meta WhatsApp with your own number – not in live operation yet
Stripe payments
Langfuse (EU) traces of phone calls for quality control, by default with the call text as spoken; kept for 30 days (Langfuse's project retention)
PostHog (EU) product analytics in the app, only with cookie consent; in addition the server sends pseudonymous sign-up and funnel events (a random visitor ID, no name, e-mail or IP address) without cookie consent
Google Analytics, Google Ads statistics and advertising measurement on the public website (not in the signed-in app), only with cookie consent
Reddit advertising measurement: pixel on the public website only with cookie consent; conversion message from the server, see above
Usercentrics cookie consent

The portals (ImmoScout24, Immowelt, Kleinanzeigen) receive only the listing data you publish, never applicant data.

Encryption

  • In transit: all web traffic uses HTTPS; plain HTTP is redirected and HSTS is set.
  • At rest: databases and storage use Google Cloud's default encryption.
  • Credentials: access data for portals, FTP and WhatsApp is additionally encrypted with AES-256-GCM.
  • Applicant data in the database is not additionally field-encrypted.

Access and roles

  • You sign in with e-mail and password (at least 8 characters) or with Google. Two-factor sign-in is not offered at present.
  • Within an account, the roles Admin, Write and Read apply (currently in the test environment only): everyone reads everything; changes need Write or Admin; payment, team and account deletion need Admin. See Organisation and team.
  • UrbanGround support can open your account to help you. In this mode changes are blocked except for viewing slots and bookings, lead comments and status, replies to applicants and the screening settings; such actions are recorded with the staff member's identity.
  • Each account's data is separated by account; every query and stored key is scoped to the account.

Audit trail

  • Changes to records are written to an append-only audit log with who made them – a user, UrbanGround staff, a service (Nimmi included) or a scheduled job – recorded as an ID, never as a name. Audit entries are deleted automatically when the retention period set for audit logs ends.
  • Lead and publishing events cannot be changed after they are written.
  • Changes to configuration are versioned.
  • There is no audit-log view in the app; ask support if you need an extract.

Retention and erasure

These deletions run automatically:

Data Deleted after
Call recordings (audio) 90 days
Documents of applicants not chosen 7–90 days after the closing message, set by the admin (default 90)
E-mail and WhatsApp enquiries without activity 365 days by default, unless a viewing is booked, a tenancy is running or a hand-over or deletion request is open
WhatsApp messages from unknown or ignored numbers 30 days by default
WhatsApp voice notes 90 days
Unconfirmed sign-ups to the interested list 7 days
Interested-list entries when the consent is withdrawn or expires (12 months without renewal)

Other data – for example call transcripts and phone-only leads – is kept until the lead is deleted. Deleting the account does not physically delete all of it; see Data export and account deletion. Records of a running tenancy are kept for the statutory period.

Deleting an applicant's data. There is no action to delete a single lead on your own initiative; data is deleted when you approve an applicant's deletion request. When you approve a request made through the unsubscribe page, contact details and message contents are anonymised, stored files of the lead are deleted and the address is blocked for further e-mail. Call records are deleted at Twilio where a Twilio call ID is stored for the call; calls Nimmi places herself through LiveKit have none. Call recordings and transcripts disappear from every view at once; the transcript text is not yet physically deleted, and the audio file is removed by the 90-day rule. Call traces at Langfuse are not deleted by this step; they expire after 30 days. Approving a deletion request from a chat runs the platform's full GDPR erasure instead, which also physically deletes the call records including transcripts and deletes the lead's traces at Langfuse; the same erasure runs when an inactive lead is purged.

Deletion requests from applicants.

  • A request in an e-mail or WhatsApp chat comes to you as a hand-over. Nimmi stops replying; nothing is deleted until you decide. You are reminded after 21 days.
  • A request through the unsubscribe page in an e-mail is approved automatically after 30 days if you have not decided.

Account deletion is described in Data export and account deletion.

  • Phone: after the greeting Nimmi says she is an AI assistant acting for the landlord – or for the company, or for the account's name shown to applicants where the account has changed the consent wording – and that the call is recorded and the details passed on, and asks for consent. Only a clear yes counts; otherwise the call ends without screening. The account can replace the opening sentence and the closing question, never the notice itself. The consent answer and its time are stored with the call.
  • Recording: the recording starts with the call, before the consent question. It is kept and stays visible to the account even if consent is refused.
  • E-mail: every e-mail from Nimmi carries an AI notice and the imprint; the account cannot remove them.
  • WhatsApp: Nimmi writes first only with consent (see WhatsApp Business). Her first reply starts with a note that she is an AI assistant, and a message denying that she is an AI is blocked. STOP, STOPP or ABMELDEN ends all WhatsApp messages.
  • Interested list: one consent for e-mail, WhatsApp and phone, valid for 12 months, with a renewal request 14 days before it ends; sign-ups through the link are confirmed by e-mail.

Anti-discrimination guardrails (AGG)

  • Nimmi never asks about ethnic or national origin, religion or belief, disability or health, age, sexual orientation, pregnancy or family planning, political views or marital status – nor about substitutes for them.
  • Allowed criteria are household size, income, move-in date, pets, profession (for the income check) and WBS. See Screening criteria.
  • Protected information an applicant volunteers is not recorded or used; a chat answer naming a protected characteristic is refused.
  • Your own screening questions are checked automatically before they are saved; questions about protected characteristics are refused.
  • Instructions and texts you give Nimmi are checked by a rule-based filter.
  • Whether an applicant meets your criteria is computed from your rules by the platform, not decided by the AI model.
  • Before an e-mail or WhatsApp message reaches the model that writes Nimmi's replies, e-mail addresses, phone numbers, IBANs, street addresses and names are masked. This masking does not apply to the classification of e-mails from new senders, the summary of a new lead or the other features in the table under Where data is processed. On the phone, the model necessarily hears the conversation.

Last updated: