← Back to Home

Privacy Policy

Last Updated: September 14, 2026

1. Introduction

Intentioned.tech ("we", "our", "the application") is a proprietary, self-hosted social skills training platform. This privacy policy explains how data is collected, stored, and used when you use this application.

Local Processing: This application is designed to be self-hosted. Your practice conversations (the audio, the transcript, and the analysis of it) are processed on the machine running the application: your own computer or server if you self-host, or ours if you use the online demo (section 2a). Your voice recordings are never sent to a third-party service.

Connections made without any configuration. Some network connections happen by default, with no setting needed to turn them on: license validation with our license service (when the server starts, every two minutes while it runs, and when a practice session ends); a daily update check on installations with the automatic updater; and downloads of AI models from their publishers, such as Hugging Face, the first time each model is needed. None of these include your audio, transcripts, or analysis results. Sections 6b and 9 describe what they send. The remaining services in section 6 are used only when chosen: subscribing to the mailing list, or an operator turning on online text-to-speech.

No cloud AI providers. The application has no option to send your conversations to a cloud AI provider such as OpenAI, Anthropic, or Google. Speech recognition, the language model, and analysis all run on the machine running the application. The only AI step that can run elsewhere is optional online text-to-speech, which is off by default (section 6e).

2a. Online Demo vs. Self-Hosted

The full application is designed to run locally on your own hardware. The online demo at app.intentioned.tech is hosted by us and is not fully local.

Online Demo Notice: When you use the online demo, your audio, transcripts, and prompts are transmitted to our servers to generate responses. Do not submit sensitive, regulated, or confidential information in the online demo. We may retain limited operational logs for security, abuse prevention, and service reliability. Your data is never used for training. Data may be deleted upon request.

"No Data Leaks" Clarification for Web Demo: The "no data leaks" claim does not apply to the web demo at app.intentioned.tech. As the owner of the demo infrastructure, Intentioned.tech necessarily has access to data processed on our servers. However, as an operator, we commit to not leaking, selling, or sharing your data with third parties. For true "no data leaks" protection, use the self-hosted version where you are both owner and operator of your data.

2. Data Collection

The application may collect and process the following types of data during your training sessions:

• Audio recordings: Voice input captured during conversations (processed locally, not stored after transcription)
• Transcribed text: Your spoken words converted to text for conversation analysis
• Conversation transcripts: The full dialogue between you and the AI assistant
• Webcam data: If enabled, this feature detects only whether a face is present in each video frame — a true/false value plus the detector's own confidence score. It does not compute facial landmarks, expressions, gaze direction, or head position, and no images or video are stored or transmitted; only that per-frame presence value and confidence score are sent to the server.
• Performance metrics: Speaking pace, response times, filler word counts, and other communication metrics

3. Safety Violation Logging

Important: Each message you send is checked by an AI moderation model running on the machine running the application. When a message is flagged, a record is created to help maintain platform safety and improve moderation accuracy.

Safety violation logs are stored on the machine running the application, in the Documents folder of the account the server runs as:

• Windows: %USERPROFILE%\Documents\simulation_safety_violations\
• macOS: ~/Documents/simulation_safety_violations/
• Linux: ~/Documents/simulation_safety_violations/

Each safety violation log contains:

• Timestamp of the incident
• A random session identifier, generated for each connection (it is not derived from your account, but the transcript below may identify you)
• The message that triggered the violation, word for word
• The reason for flagging
• Full conversation transcript up to that point

Purpose: These logs are stored locally for your own review and to help improve the moderation system.

Reports for the host/administrator: A second AI check decides whether each flagged message repeats an earlier flagged message in the same session, and rates its severity as severe, moderate, or mild. When a session reaches its third distinct flagged message, and each time a message is rated severe, a report containing the flagged messages is written to a separate "transmitted_to_host" folder for the host/administrator to review. Despite the folder's name, nothing is sent over the network; the report stays on the same machine, in [Documents]/simulation_safety_violations/transmitted_to_host/

Pausing a conversation: From the fifth distinct flagged message in a session, or once messages rated severe reach two, the conversation is paused: a pause notice is shown and flagged messages are no longer answered. Because each flagged message is currently rated twice, a single message rated severe can be enough. Reloading the page starts a new count. If a crisis signal was detected earlier in the conversation, the pause is not applied, so that crisis resources stay reachable. These thresholds are fixed; the operator of an instance can turn content moderation off or change what it flags.

How long logs are kept: When automatic deletion is on, which it is by default, logs and reports older than the retention period (30 days by default) are deleted. This check runs when the server starts, so on a server that runs for a long time without restarting, a log can be kept longer than the retention period.

3a. Safety Notices

Every session shows a notice that you are talking to an AI, not a human, when the session starts and again after each hour of practice. If something you say matches the application's crisis detection, which currently recognizes English only, crisis resources are shown, including the 988 Suicide and Crisis Lifeline. No setting turns either notice off.

4. Cookies and Local Storage

This application uses cookies to remember your preferences across sessions:

• Settings Cookies: Your TTS/STT engine preferences, voice selections, LLM parameters
• UI Preferences: Selected scenario, microphone mode, mute state

Cookie Details:

• Cookies are stored locally in your browser only
• No cookies are transmitted to external servers
• Cookies expire after 1 year of inactivity
• You can clear cookies at any time through your browser settings

Cookies used by this application:

• intentioned_tts_engine - Your preferred text-to-speech engine
• intentioned_kokoro_voice - Selected Kokoro TTS voice
• intentioned_vibevoice_voice - Selected VibeVoice voice
• intentioned_stt_engine - Your preferred speech-to-text engine
• intentioned_max_tokens - LLM response length setting
• intentioned_temperature - LLM creativity setting
• intentioned_scenario - Last selected training scenario
• intentioned_mic_mode - Preferred microphone mode (VAD or Push-to-Talk)
• intentioned_tts_muted - Whether AI voice is muted

5. Data Storage

Everything in this section stays on the machine running the application: yours, if you self-host; ours, if you use the online demo. Live conversation audio is never written to disk; it is transcribed in memory and discarded. The conversation transcript is held in memory during a session and is not saved on its own, except where listed below. The following is written to disk:

• Session results: each time a session is analyzed, a file in session_results/ records the scores, the feedback and summary generated from your conversation, the scenario, a session identifier, and your username. It does not contain the verbatim transcript or any audio. Results are deleted automatically after the retention period (30 days by default).
• Track B history: if you opt in to Track B, the same record is also kept under your username in cold_storage/, indefinitely (section 6b-1).
• Safety violation logs and reports: your flagged messages, word for word, and the conversation transcript up to each flag (section 3).
• Accounts and consent: for each account, the username, a hash of the password, the erasure PIN, whether and when you opted in to Track B, when onboarding was completed, the policy version accepted, and the account's role. Each consent record stores the time, the policy version, the type of consent, and an IP address. The account and erasure-PIN files keep their five most recent previous versions as backups, so a deleted account can remain in a backup until five later changes have replaced it.
• Safety notice timing: so that the notices in section 3a appear at the right times, the server records, for each browser, how much practice time has accrued, when it was last seen, and whether a notice was acknowledged. These records are keyed by a hash of a random identifier your browser keeps in local storage (intnd_subject_id), not by your account, and are removed after 30 days without use. A record that the one-time suitability notice has been shown is kept indefinitely.
• Safety counters: yearly totals of how often each safety protection acted: AI output blocked by the self-harm or sexual-content checks, crisis resources shown, questions about whether you are talking to a person, and notices that could not be confirmed. The totals contain no identifiers or message content and are kept indefinitely.
• Safeguarding audit log (off by default): if the operator turns it on, one line per detected crisis signal with the time, a shortened session identifier, the category, whether a contact was notified, and optionally the username. It never contains the message itself.
• Operator and license files: the application's configuration, operator and demo login credentials, and license activation data (section 9). These contain no conversation content.
• Server log: the server writes a running log to the host's system journal. It includes the AI's replies in full, short excerpts of AI output that a safety check blocked, usernames in data-retention and erasure messages, and the IP address and path of each web request. What you say is included only if the operator turns on transcript logging, which is off by default. The log is kept according to the host's journal settings; the Linux installer limits it to 7 days by default.
• Temporary files: audio you upload for file analysis is written to a temporary file and deleted when the analysis finishes. When online text-to-speech is in use, the synthesized AI speech is briefly written to a temporary file; if synthesis fails, that file can be left behind.
• Model files: downloaded AI models are cached on disk. They contain no user data.

6. Third-Party Services

The application may interact with external services for the following purposes:

• License validation and update checks: our license service, hosted on Cloudflare (section 9)
• Model downloads: AI model publishers such as Hugging Face, the first time each model is needed (section 6b)
• Online text-to-speech (off by default): Microsoft, only if the operator turns it on (section 6e)
• Email delivery (Resend): When you subscribe to our mailing list, we use Resend to send transactional emails (welcome emails, announcements, unsubscribe confirmations). Resend processes your email address to deliver these messages. See Resend's Privacy Policy for details on their data handling practices.

The face-detection library used by the optional eye-contact feature (face-api.js) and its model files are served by the application itself. Nothing is loaded from a CDN, and the application's content security policy does not allow scripts to be loaded from other sites.

Three of the four text-to-speech engines (Kokoro, VibeVoice, pyttsx3) run completely offline on the machine running the application. The fourth, Edge-TTS, does not; see section 6e.

6a. Mailing List

The mailing list is a separate, optional service run from this website. It is not part of the Intentioned.tech application: subscribing to it does not connect to your practice sessions, transcripts, or analysis results in any way, and nothing from the application is ever added to it.

What we collect: your email address, and nothing else. The signup form sends only the address you type into it. We do not collect your name, your browser's language, your IP address, or any other field at signup.

Where it is stored, and in which country. Your address is stored by Resend, our email provider, as a contact in our mailing list. Resend is where the list actually lives — it is not merely a delivery pipe — so your address is held on Resend's systems, subject to Resend's privacy policy. We do not keep a separate copy of the list in a database of our own.

Those servers are in the United States. We are telling you this plainly because it is the kind of thing people assume is configurable and it is not: there is no European storage option, and we have not chosen the United States over an alternative. If you are in the UK, the EU, or Switzerland, your address is transferred to the United States under the data processing agreement we have in place with Resend, which is the legal mechanism that permits the transfer.

Why we are allowed to use it. We rely on your consent, given when you submit the signup form. You can withdraw that consent at any time, and withdrawing it is as easy as giving it — one click on the unsubscribe link in any email we send.

How long we keep it. If you unsubscribe, we stop mailing you immediately and delete your address within 30 days. We hold it for those 30 days and no longer, for two reasons: an unsubscribe clicked by accident can be undone, and a stale backup restored during that window cannot quietly put you back on the list. After 30 days nothing about you remains.

If you never unsubscribe, we keep your address until the launch you signed up for, and in no case longer than 24 months from the day you signed up. If we have not launched by then, we delete the list rather than email people about something they asked to hear about two years earlier and have every right to have forgotten. You can ask us to delete you sooner at any time, and we will.

Access and deletion. Write to contact@intentioned.tech and we will tell you what we hold for your address, or delete it, within 30 days. There is no automated self-service portal for this, and we are not pretending otherwise: it is a person reading email at a three-person company.

6b. Clarification: "No Data Leaks" and "No Cloud"

What These Claims Mean:

• "No data leaks" refers specifically to your conversation data, audio recordings, and transcripts processed by the Intentioned.tech application on our infrastructure. This type of collection of data does not happen within the application, so we cannot possibly have a data leak from it in this regard. It does not apply to: (a) your local machine security, or (b) third-party LLMs you configure, or (c) payment processing or other interactions with our site, or (d) the online demo.
• "No cloud" and "no latency" refer to the core conversation interface. External services (payment processing, mailing list, license validation) use cloud infrastructure. Downloading models directly from the application may use the internet. "No Latency" refers to zero network latency via WAN.
• Versions that are unofficial, are known to be compromised, have known vulnerabilities, are paired to the cloud, or have been shipped via a malicious installer via an outside threat actor are not covered by the "No Data Leaks" claim. We will make best effort mitigations in these scenarios according to our judgement.
• While conversation processing is entirely offline, downloading the initial AI models requires a connection to third-party repositories (e.g., Hugging Face). Your IP address and download request will be subject to their privacy policies during the download phase.
• All conversation data remains on your local machine. We do not apply proprietary encryption to these local files; securing this data at rest is dependent on your operating system's security (e.g., using BitLocker, LUKS, or FileVault).
• Track B (AI Analytics / indefinite retention) is excluded from this guarantee. If you opt in to Track B, the data you choose to place under it is retained indefinitely and processed for AI analytics. The "no data leaks" guarantee does not apply to Track B data. See section 6b-1 below.

6b-1. Track B Analytics (Optional AI Analytics & Indefinite Retention)

By default, your data is handled under Track A and is covered in full by the "No Data Leaks" guarantee described above. Track B is a separate, optional feature you must explicitly opt in to.

• What it does: Track B retains your derived session metrics and AI-generated summaries indefinitely and processes them for personal AI analytics, so you can review past performance and receive enhanced AI feedback across multiple sessions. It does not retain your verbatim conversation transcripts — the same non-transcript record Track A already produces (scores, short feedback text, and a generated summary), just kept indefinitely instead of being automatically deleted.
• Retention: Track B data is kept indefinitely — it is not subject to the automatic 30-day deletion that applies to Track A — until you withdraw your consent or delete it.
• Processing: Track B data is analyzed by an AI model to produce historical insights and aggregated performance feedback.
• Exclusion from the guarantee: The "no data leaks" guarantee does not extend to data you choose to place under Track B — indefinite retention and ongoing AI processing carry different risk than Track A's automatic deletion. This does not mean Intentioned.tech obtains your data: Track B is processed locally, the same as Track A, and Intentioned.tech does not obtain it. By opting in, you are trading Track A's stronger retention guarantee for the indefinite-retention analytics feature — not the separate assurance that Intentioned.tech will not obtain your data, which still holds.
• Withdrawing consent: You may withdraw Track B consent at any time. After withdrawal, no new data is added under Track B; you may also delete previously retained Track B data at any time.

6c. External LLM Providers

The application does not support cloud-based Large Language Model (LLM) providers. No setting sends your conversation to OpenAI, Anthropic, Google, or any other hosted AI service: the language models it uses run on the machine running the application.

• You are responsible for adhering to the specific licenses of the third-party models you choose to download and run. Intentioned.tech provides the infrastructure, but does not grant you the license to the underlying model weights.

6d. Payment Processing & External Services

The following data may be collected and stored outside the local application for service functionality:

• Payment information: Processed by third-party payment providers (e.g., Stripe, PayPal). We do not store full credit card numbers.
• License activation: your license key or account name, server fingerprints and hostnames, and activation and check-in times, stored in our license service's database on Cloudflare (section 9)
• Mailing list: Email addresses as described in section 6a

How this data is handled:

• Deletion on request: write to contact@intentioned.tech and we will delete the license, account, and mailing-list data we hold about you. Deleting a server's activation record deactivates that server.
• Mailing list transfers: your address is held by Resend in the United States under our data processing agreement with Resend, which is the transfer mechanism for subscribers in the UK, the EU, and Switzerland (section 6a)
• Collected only where necessary for service functionality (licensing, communications, payments)
• Never shared with third parties except as required for service operation

6e. Online Text-to-Speech (Edge-TTS)

The application can optionally use Microsoft's Edge-TTS service instead of an offline engine. When it is used, the text of the AI character's spoken reply is sent to Microsoft's speech service (speech.platform.bing.com) along with a voice identifier (e.g. "en-US-JennyNeural") that encodes the selected language and voice. Your own audio and anything you typed are never sent, although the AI's reply can repeat or paraphrase what you said. No account identifier, session identifier, or other identifying data from this application is included; as with any network request, the server's IP address is visible to Microsoft as the connecting party.

Off by default. Online TTS is disabled unless the operator of the instance turns it on: in the server's configuration, with an environment variable set before the server starts, or through a settings change that requires operator credentials. Ordinary signed-in users cannot turn it on. It is currently off on the online demo.

What triggers it. Once online TTS is turned on, it is used in three cases:

• The TTS engine is set to Edge-TTS, in which case it is used for every reply, in every language.
• The reply is in a language other than English, whichever engine is selected: every non-English reply is sent to Edge-TTS while online TTS is on. Non-English audio is not exempt from this; if anything, it is more likely to use it than English audio is.
• The TTS engine is set to Kokoro and Kokoro's synthesis fails for any reason, in which case Edge-TTS is used as a fallback.

The application's own safety notices, such as the AI disclosure and crisis resources, are always synthesized offline.

No active notice when in use. The application does not currently display anything to a user indicating that a given reply's audio was generated online rather than offline. A tooltip describing this exists in the application's source code, but the control it is attached to is hidden in the current interface, so no user currently sees it.

7. Browser Extension Detection

Malwarebytes Browser Guard Detection: This application includes code to detect when Malwarebytes Browser Guard is flagging our domain as a "Risky TLD" (false positive). This detection is used solely to display a helpful warning banner explaining how to whitelist the site.

The detection works by:

• Attempting to load a small invisible resource from our own domain to detect if it's being blocked
• Monitoring console logs for Browser Guard-specific messages
• Checking for DOM elements injected by the extension

This detection is entirely local and does not contact any external Malwarebytes servers.

No data is collected or transmitted through this detection. It exists purely to improve user experience by explaining false positive blocks on .tech domains.

8. AI Models Used

This application uses the following AI models, which run locally on the machine running the application:

• Speech-to-Text (STT):
  • NVIDIA Parakeet TDT 0.6B v3 (default) - NVIDIA NeMo, CC-BY-4.0 license, 600M parameters, supports 25 languages
  • Vosk - Apache 2.0 license, offline-first, lightweight, 20+ language models
• Text-to-Speech (TTS):
  • Kokoro-82M (default) - Apache 2.0 license, 82M parameters, StyleTTS2 architecture
  • VibeVoice Realtime 0.5B - 500M parameters, higher quality voice cloning
  • pyttsx3 - Local system voices, fully offline fallback
• Language Model (LLM):
  • Qwen2.5-3B-Instruct (default) - 4-bit quantized, ~2GB VRAM
  • Language-specific models available for 28 EU languages; safety checks are English-only (section 13)
• Eye Contact Detection: Face-API.js - browser-based, no images stored

All AI processing occurs locally, except optional online text-to-speech (section 6e). Model weights are downloaded once and cached on the machine running the application.

9. License Key System

License Validation: A licensed installation of Intentioned.tech communicates with our license service, which runs on Cloudflare, to confirm that its license is valid. It does this when the server starts, every two minutes while the server runs, and when each practice session ends. Installations with the automatic updater also check for updates once a day.

The license service receives:

• License key or account: your license key, or your account name and the activation token issued to your server. When you activate a server with an account, your account password is also sent, over an encrypted connection.
• Server fingerprint: a hash of the server's hostname, network (MAC) address, processor description, and operating system version, used to count your servers against your license's activation limit (3 servers unless your license specifies otherwise)
• Server hostname: sent in readable form when a server is activated
• Session identifier: a random identifier for each practice session, which may be sent when the session starts
• Version: the installed version, with each update check
• Activation times: when each server was activated, and when it last checked in

IP address: as with any network request, your server's IP address is visible to Cloudflare, which hosts the license service. The license service does not record it in its database.

What is NOT collected:

• No conversation data, audio, or transcripts
• Nothing about what you practice or what you say. The service does learn when practice sessions end on a licensed server, because it is notified then.

Activation records are kept until the server is deactivated, or until you ask us to delete them (section 6d).

10. Data Retention

• Conversation audio and transcripts: held in memory during a session and cleared when it ends, except where recorded in safety violation logs (section 3)
• Session results: deleted after the retention period, 30 days by default (section 5)
• Safety violation logs and reports: deleted after the retention period, 30 days by default, checked when the server starts (section 3)
• License activation data: retained by the license service until the server is deactivated, or until you ask us to delete it (section 9)
• Mailing list subscriptions: if you unsubscribe, your address is deleted within 30 days; if you never unsubscribe, it is kept until the launch you signed up for and no longer than 24 months after you signed up. See section 6a

11. Your Rights

Since this is a self-hosted application, you have complete control over your data:

• Access: All data is stored locally and accessible to you
• Deletion: You can delete any stored data at any time
• Portability: Session results are stored as CSV files and other records as JSON, for easy export

The mailing list and license records are the exceptions, because we hold them rather than you: they are not on your machine. Your access and deletion rights over them are described in sections 6a and 6d, and are exercised by writing to contact@intentioned.tech.

12. Security

Data security depends on your deployment configuration. We recommend:

• Using HTTPS for all connections
• Keeping your system and dependencies updated
• Restricting network access to trusted users only
• Regularly reviewing and clearing safety violation logs

13. Multilingual Support

Intentioned.tech supports 28 EU languages. Language detection and preferences are:

• Stored locally in your browser's localStorage
• Used only to select appropriate AI models and UI translations
• Not transmitted to external services, except as part of the voice identifier sent when online text-to-speech is in use (section 6e)

Safety features are English-only. Crisis detection (section 3a) and the checks that block harmful AI output recognize English text only. In a conversation held in any other language, crisis detection does not fire and those checks will generally not match. The application is not restricted to English by default; the operator of an instance can restrict it to a single language. The hosted instance at app.intentioned.tech is restricted to English.

14. Changes to This Policy

This privacy policy may be updated as the application evolves. Check the "Last Updated" date at the top of this document for the most recent version.

15. Contact

For questions about this privacy policy or the application, please contact us at contact@intentioned.tech or visit intentioned.tech.