Legal
Keeni AI Privacy Policy
Effective: · Version: privacy-2026-10-28
The short version
- You create an account before a mock. We record your voice for IELTS Speaking practice and give you a score and report.
- Your scores are AI practice estimates, not official IELTS results.
- We use scores, measurements, timings and errors from mocks to make Keeni more accurate. Optional research using learner recordings and transcripts is currently unavailable. You can still withdraw historical research permission.
- We don't sell your data, show ads, or use your voice to identify you.
- You can delete any recording, any mock, or your whole account.
1. Who we are
Keeni AI (keeni.ai) is run by Lean Start Lab LLC, doing business as Keeni AI, a company in Missouri, United States ("Keeni", "we", "us"). We decide how your data is used, so we are responsible for it. Questions or requests: privacy@keeni.ai.
2. What we collect
| What | Examples |
|---|---|
| Account | Email address, optional display name, birth year (and month if your year is on an age boundary), country |
| Practice | Your mock recordings, transcripts, pronunciation and fluency measurements, scores, reports, and microphone-check audio |
| Your choices | What you agreed to, when, and under which version of this policy |
| Device and usage | Steps used, timings, errors, browser and device information, microphone and connection quality, country inferred from the network, and pseudonymous identifiers such as hashes of an IP address or device ID |
| Messages | Emails you send us |
Hashes and internal IDs can still link records to a person or mock. They do not make all usage data anonymous. Cloudflare also receives browser and network signals for the bot check (section 5).
3. How we use it
To run Keeni for you. Check your microphone, run your mock, decide when an answer has finished, transcribe and score it, write your report, let you replay recordings you chose to keep, send sign-in codes, secure your account, stop bots and abuse, and enforce free-use limits.
To make Keeni better. We use scores, measurements, timings, error reports and usage events to find bugs, check that scoring is accurate and fair across different speakers, and improve the examiner and the scoring. If you report a problem, or our checks flag a mock as possibly wrong, a small number of people at Keeni may look at or listen to that mock to fix it.
Optional research participation is currently unavailable. We are not accepting new participation or using learner recordings and transcripts for optional research while it is unavailable. Your ordinary mock, score and report remain available.
Research permission (Consent A) covers study, training and testing of Keeni's own scoring models, including comparisons of speech services. It covers only mocks started while the applicable permission was on, within the scope of the words you agreed to. It does not change your original score or give us permission to publish or share your recordings, transcripts or quotes from your answers. Research permission starts off; saying no does not change your score or access to the mock. You can still withdraw historical permission in Privacy settings or by email. The existing research deadlines in sections 9 and 10 continue to apply.
We also use combined statistics to check service quality. Removing a name alone does not make a recording, transcript or linked measurement anonymous.
4. How scoring works
Your examiner's voice is AI-generated. When you finish a mock:
- Speech-to-text turns your recording into a transcript with timings.
- A listening model receives the learner-answer audio from the whole mock to assess pronunciation.
- Text judges receive your transcript, the questions, timing measurements and a Soniox-derived timing/coverage summary to assess fluency, vocabulary and grammar against the public IELTS Speaking band descriptors. The text judges do not receive audio in those requests.
- The server combines the criterion bands into an overall practice band and report.
These are practice estimates. They do not determine admission or a visa. If a score looks wrong, contact us for review.
5. Who we share it with
These services process data for the functions below. Conditional processing depends on the services used in a deployment; the table does not mean every conditional function is active. Provider processing also depends on applicable terms and settings. We do not use your voice to identify you, create voiceprints or clone your voice.
| Company | What they do | What they get |
|---|---|---|
| Supabase | Database and authentication; recording storage where that service is used | Account, transcripts, scores, reports and choices; encrypted recording objects when used for storage |
| Railway | Runs our servers | Data passing through the app |
| Cloudflare (Turnstile) | Bot check | IP address, browser signals and site/origin information. Cloudflare acts as processor for the site-protection service and as controller when improving Turnstile's bot detection, under its Turnstile Privacy Addendum |
| Soniox | Live transcription during the mock; asynchronous microphone check and scoring transcription | Your voice, producing recognised words and timings |
| OpenRouter, routing to Google (Gemini) | Whole-mock pronunciation listening | Learner-answer audio |
| OpenRouter, routing to Google (Gemini) | Text judges for fluency, vocabulary and grammar | Transcript, questions, timings and Soniox-derived timing/coverage summary; no audio in these requests |
| OpenRouter, using the TypeSafe Jev decision model | Detecting whether an answer has finished | Current confirmed and provisional answer words, question and timing context; no audio in this request |
| OpenRouter and the model provider selected for coaching | If configured, prepares answer rewrites; this processing can run even when coaching is not shown in the app | Selected answer text, question, scoring criterion, band, target band and improvement guidance; no audio in this request |
| Microsoft Azure | Speech comparisons for optional research, which is currently unavailable; outside ordinary mock scoring | Recordings and reference text only where covered by applicable research permission |
| Inworld | Synthesising the examiner's voice | Examiner question and instruction text for the current scripted room |
| Resend | Sends internal incident alerts if email alerting is configured | Alert subject and text, operational identifiers, error context and the sender and recipient service email addresses |
Soniox distinguishes live processing from stored asynchronous jobs. Its security and privacy documentation says audio and transcripts are not used to train its models, and live processing does not retain them unless a storage feature is requested. Stored async files and transcripts are automatically deleted after 30 days, unless deleted earlier. We request earlier deletion of our async files; if that cleanup fails, the automatic 30-day deletion remains the fallback. This is not a promise of immediate removal after every request.
For model calls, no-training routing and zero retention are different controls. OpenRouter's ZDR documentation distinguishes them. Its support notice describes API metadata logging and optional prompt/completion logging; its privacy policy describes security/legal retention exceptions and differing model-provider practices. We do not promise that every recipient keeps nothing or cannot use data for another purpose.
We may also disclose data if the law requires it, to protect someone's safety or our legal rights, or if Keeni is sold or merged (the new owner keeps this policy's promises).
6. Where your data is processed
Our hosting, speech and model providers may process data outside your country, in countries where the services used for your mock operate. We do not promise that all processing or storage stays in one country. Contact us for information about the processing of your data.
7. Cookies and analytics
We use sign-in cookies and a random device cookie for session, free-use and age-check controls. Our records use hashed device identifiers. Our own usage records can include user or mock IDs; these records are not all anonymous. Usage analytics events have a 90-day retention period. Timings stored with a mock follow that mock's retention. Cloudflare receives bot-check signals as described above.
8. Age
Keeni is for people 18 and older right now. After sign-up, we ask for your birth year, and month where needed, before recording. The adults-only gate blocks under-18s. Age declarations include the answer, date, country and hashed device identifier. If you think someone under 18 has used Keeni, email privacy@keeni.ai.
9. How long we keep it
| Data | How long |
|---|---|
| Ordinary recordings, replay OFF and research OFF | Kept for processing and any available answer retry. New answer retries close 7 days after the final report. An already opened retry can retain needed audio while it finishes. Once that window has closed, an unsubmitted retry that has been open for at least a day is eligible for release. Eligible audio is then queued for deletion; this is not a promise that every byte disappears at day 7 |
| Ordinary recordings you choose to keep | Until deletion, a selected auto-delete period (30, 90 or 365 days), or 365 days of account inactivity, whichever applies first |
| Research use of recordings and transcripts, and research copies including exports and backups | At most 24 months from the original mock recording, or earlier withdrawal/deletion. Exporting, re-granting permission or labelling a mock does not restart this limit |
| Transcripts, measurements, scores, reports | While your account is open. Deleting a recording keeps its transcript, score and report |
| Usage analytics events | 90 days. Timings and events that are part of a mock stay until the mock or account is deleted |
| Operational logs and backups | Limited periods for service operation, recovery and security, subject to legal requirements. Research copies in backups remain subject to the research limits in this policy |
| After you delete your account | Limited consent, deletion, security and free-use records, including hashed identifiers, may remain to prove choices and prevent abuse; age declarations linked to your account are deleted with it |
"Keep my recordings" starts OFF for a new learner. A previously saved replay choice is shown when you return. Replay OFF blocks learner replay. Optional research participation is currently unavailable. This pause is not a withdrawal and does not extend or restart existing research deadlines. At the research deadline, research use ends and research copies must be removed. Withdrawal or deletion of a covered mock or account requires removal of its research copies within 15 days.
Your original service recordings, transcripts, scores and reports follow their ordinary settings and deletion rules. Ending research retention or withdrawing research permission does not by itself erase them or change your score. They remain unavailable for research while participation is unavailable, after withdrawal or after the research deadline.
10. Your rights
You can:
- see your mocks, reports and history in the app, and ask us by email for a copy of your data
- correct your details
- request deletion of a recording, mock or account through the app's privacy controls
- withdraw any permission at any time (this stops future use; it doesn't undo past use)
- object to a use, or ask us to limit it
- complain to the data protection authority where you live
Optional research participation is currently unavailable. You can still withdraw historical research permission in Privacy settings or email privacy@keeni.ai. We handle research withdrawal within 15 days, including stopping research use and removing matching research copies. The pause does not count as your withdrawal or extend this deadline. Other rights requests follow the deadlines required by the applicable law; we tell you if a lawful extension is needed and why.
Withdrawal stops future research and training use of the covered recordings and transcripts. It does not change your original score or automatically rebuild models and research results already produced.
11. Security
Stored learner audio is encrypted with a separate key for each file. The app checks ownership before returning private recordings and reports. Authorised systems, service providers and staff may process data for the purposes described here. No system is perfectly secure. If a breach affects you, we'll tell you and the authorities as the law requires.
12. Changes
If we change this policy in a way that matters, we'll tell you by email or in the app before it takes effect. We won't use your data for a new purpose that needs your permission without asking first.
13. Contact
Lean Start Lab LLC, doing business as Keeni AI, Missouri, United States. Email: privacy@keeni.ai.