Privacy Policy for Owli-AI Assist

3.1 Image data

When you trigger a scene request, the app processes an image or snapshot of the camera scene.

The following applies:

• The pure live preview runs primarily on the device itself.
• Data is only transmitted externally once you actively use a corresponding AI function, especially "New scene", a follow-up question, or a similar intentionally triggered AI request.
• In backend mode, the snapshot is transmitted to the Owli backend and from there forwarded to OpenRouter or model providers connected there for AI evaluation.
• In backend mode, follow-up questions additionally transmit the original scene image together with the question text to the Owli backend, and from there again to the VLM provider, so that the AI can answer with image context.
• The original scene image is not stored permanently in the Owli backend for follow-up questions and is not embedded in the sceneToken.
• In BYOK mode, the snapshot is transmitted directly to OpenRouter.
• In BYOK mode, if you actively use corresponding features, the original scene image, optional additional image attachments, and your question text may be transmitted directly to OpenRouter.

Depending on what is captured, image data may also contain sensitive information such as faces, documents, addresses, health information, or other private content. Please do not use the app to process highly sensitive content unless this is necessary.

3.2 Text and voice input

If you ask a follow-up question, the question text is processed:

• in backend mode via the Owli backend; in current app versions, the original scene image is additionally transmitted again so that the follow-up can be answered with image context,
• in BYOK mode directly via OpenRouter.

If you use voice input, for example to ask a follow-up question via microphone, the app or the speech recognition service available on your device may convert your speech into text. Depending on the Android version, device settings, and the selected speech recognition service, this processing may take place locally on the device or by the respective provider of the speech recognition service.

If the app provides a live mode or direct voice conversation in the future and you actively use that feature, voice data, derived text, and the current multimodal context may be processed in order to provide the ongoing conversation. Such processing only takes place while you actively use the corresponding feature.

3.3 AI answers and reported content

The app receives AI answers back as text. These answers are displayed in the app and can optionally be read aloud via text-to-speech.

If you use the in-app function "Report AI answer", the app additionally processes data relating to a report of a problematic AI answer. This may include in particular:

• reported answer text,
• selected report category,
• optional comment,
• optional associated question text,
• app version and version number,
• language / locale,
• selected transport mode,
• profile ID,
• installation identifier or pseudonymized installation correlation,
• existing request ID / answer ID,
• time of the report.

AI reports do not normally include image data, screenshots, OpenRouter keys, or auth headers.

3.4 Technical metadata

The following technical data may be processed to provide, stabilize, and secure the service:

• app version and version number,
• platform (Android),
• language / locale,
• pseudonymous installation identifier,
• selected transport mode,
• profile ID,
• technical error and status information,
• request ID / answer ID,
• IP address or connection data at server/proxy level, if any.

3.5 Data stored locally

In particular, the app stores locally:

• settings such as transport mode, language, TTS options, and profile selection,
• cached public profile metadata,
• optionally, an OpenRouter key stored by the user in BYOK mode.

According to the current technical state, the BYOK key is stored locally in encrypted form.

4.1 Backend mode

In backend mode, the app processes AI requests by default via the Owli backend.

Typical process for a new scene:

1. You actively trigger a scene request in the app, for example via "New scene".
2. The app sends the required snapshot image and technical metadata to the Owli backend.
3. The Owli backend processes the request further and forwards it to OpenRouter or model providers connected there for AI evaluation.
4. The answer is returned to the app and displayed there or optionally read aloud.

Typical process for a follow-up question:

1. You actively ask a follow-up question in the app, for example via text input or microphone.
2. The app sends the question text and the original scene image again to the Owli backend.
3. The Owli backend forwards the question and image context for AI evaluation.
4. The answer is returned to the app and displayed there or optionally read aloud.

The Owli backend does not permanently store image data transmitted with AI requests as a user-related image history. The sceneToken contains textual scene context, but no image data.

4.2 BYOK mode

In BYOK mode, the app uses your own OpenRouter access directly.

Typical process:

1. You store your own OpenRouter key locally in the app.
2. You trigger a scene request or follow-up question.
3. The app sends the request directly to OpenRouter.

In BYOK mode, OpenRouter is the direct recipient of the AI data transmitted in this mode. Depending on the feature used, scene image, optional additional image attachments, and question text may be sent directly to OpenRouter.

4.3 Reporting problematic AI answers

If you use the in-app function "Report AI answer", the app sends the report data to the Owli backend. The backend validates these reports and stores validated reports in a privacy-conscious way for later review.

Stored data may include category, reported answer text, optional comment, optional question text, technical metadata, and a pseudonymized installation correlation.

The reporting function is not a fully automated moderation, triage, or enforcement pipeline. Its purpose is to mark problematic AI-generated answers in a traceable way and make later review possible.

4.4 Voice input and live features

If you use voice input, your speech is processed for recognition and conversion into text. The exact processing may depend on the Android or third-party speech recognition service used on your device.

If a live mode or direct voice conversation is available in the app and you actively use that feature, voice data, image/scene context, and derived text content may be processed continuously in order to enable the conversation. Such processing only takes place while you use the corresponding feature.

9.1 Locally on the device

Settings and caches stored locally generally remain on the device until they are

• changed or deleted by you,
• overwritten by the app,
• or removed by uninstalling the app.

9.2 Backend / server / logs

The Owli backend is not designed to create user accounts, persistent user profiles, or permanently retrievable content histories.

Image content, prompts, model answers, and TTS texts are only processed in the backend to the extent necessary for the respective request. They are not intended to be stored permanently as user-related content histories unless this is necessary for operation.

However, technical metadata and operational logs may be processed for a limited time, in particular for:

• security,
• troubleshooting,
• abuse prevention,
• stability analysis,
• load and operations analysis.

9.3 Reports of problematic AI answers

Validated reports of problematic AI answers are stored server-side for review purposes. The installation identifier is hashed in a pseudonymized form.

The standard retention period for such reports is currently 30 days. Reports do not normally include image data or screenshots.

9.4 BYOK

In BYOK mode, additional storage operations may occur at OpenRouter or service providers integrated there. Our influence over this is limited to what is determined by our technical integration.

9.5 Deletion and local control

You can delete locally stored settings by resetting app data or uninstalling the app. A stored BYOK key can be deleted in the app via key management.

Since the app does not currently provide an Owli-AI user account and the backend is not designed to create permanently user-related content histories, server-side data cannot always be assigned to a specific person or app installation. If you contact us regarding access or deletion, we will check whether personal data relating to your request exists and can be assigned to you. This may include technical metadata or AI reports triggered by you. Where such data exists and no statutory or legitimate retention reasons prevent deletion, it will be deleted.