Privacy Policy

1. Overview

The offrec api console is a visual platform for building, training, exporting, and deploying machine learning models. This policy explains what information we collect when you use the platform, why we collect it, and what controls you have. It applies to console.offrec.ai, offrec chat at chat.offrec.ai, this site, and related services operated by PnotP, Inc. d/b/a offrec.

2. Data We Collect

We collect only what is necessary to run the product:

a) Account data

• Name and email when you sign up or connect an identity provider such as GitHub
• Organization, if you provide one

b) Project & model data

• Project metadata: names, descriptions, class labels, pipeline and architecture configurations
• Training run records and resulting metrics
• Dataset files and assets you choose to register with a project

c) Usage & product analytics

• Operational logs (request paths, timestamps, error traces) for running and debugging the service
• Anonymous and account-linked usage events (pages viewed, features used such as creating a project or starting a training run) and session recordings, used to understand how the product is used and improve it

d) Inference request metadata

For every call to our inference API we keep a record of the call, and none of what was said in it:

• Which API key made it (stored as a one-way hash), when, which endpoint, which model you asked for and which one answered, which provider served it, how many tokens went in and out, how long it took, and whether it succeeded. If we declined the request, the reason code.
• Not your prompt, and not the response. The record has no field that could hold them — not a truncated version, not a redacted one. An automated check fails our build if anyone adds one. We keep this because when a model provider asks us to account for suspected misuse, this is what lets us identify which customer’s key was involved and cut off that key, without keeping what anybody wrote.

e) Billing data

• Subscription and payment information processed by our payment provider (Stripe). We never see or store full card numbers.

f) offrec chat

• We store your conversations, the files you attach and your projects for your account, so you can open them on any device you sign in on and later messages can use your files. Your browser encrypts each one (AES-256-GCM, with a key for your account alone) before sending it, and we store only the encrypted copy.
• This is encryption at rest, not end-to-end encryption. offrec holds your account’s key and releases it only to your signed-in session, so our systems can decrypt what is stored. We never use the stored copy to train models, never sell it, and never share it, and we access it only to run the service or where the law requires it. What leaves offrec is described in the next two points: each message to the model provider that answers it, and, only when you turn them on, searches and connector requests.
• Each message goes to the model through our zero-retention inference (section 5). The copy we store is separate from the model call.
• Word, Excel and PowerPoint files are turned into text in your browser. That text is what we send and store, not the original file.
• Web search and connectors run only when you turn them on. A web search sends the search query, not your conversation, to Linkup (or Ceramic.ai, if Linkup fails) through Cloudflare’s Web Search API. Both engines are listed by Cloudflare as zero data retention, and our search gateway keeps no logs of the queries. The engine used is named on each search. A connector sends requests and their results through Composio to the app you connected, and every action that changes something in that app waits for your approval.
• You can delete everything stored for your account at any time in Settings → Privacy. Deleting your account deletes it too.

3. How We Use Data

• To provide the training, evaluation, export, and deployment features you use in the product
• To keep your offrec chat conversations, files and projects available on the devices you sign in on
• To authenticate you, manage your subscription, and prevent abuse
• To understand product usage and improve performance and reliability
• To send transactional emails (password reset, receipts, important account changes). We do not send marketing emails unless you opt in.

4. Data Ownership

You retain ownership of your projects, datasets, and trained models. You can export your project and training artifacts at any time through the in-app export feature. We do not use the datasets you upload to train our own models, and we do not share them with third parties.

5. Data Sharing

We do not sell your data. We share data only with the service providers we depend on to run offrec:

• Supabase (authentication and database)
• Render (application hosting)
• Cloudflare (content delivery and static hosting)
• Brevo (transactional email)
• Stripe (payments)
• PostHog (product analytics and session recording)
• Vibiz (website analytics and ad attribution on offrec.ai; no contact details without your consent)
• Composio (offrec chat connectors, only for the apps you connect and only while they are connected) — under its zero data retention option: the requests and their results are not stored. Composio keeps the sign-in for each app you connect, so it can act for you, until you disconnect it.
• Cloudflare’s Web Search API, with Linkup or Ceramic.ai as the engine (offrec chat web search, only when you turn it on) — under zero data retention: the query is not stored, and our search gateway keeps no logs
• Fireworks AI (LLM inference) — under a zero-retention agreement: your prompts and completions are not stored at all, including for abuse monitoring.
• Google Cloud Vertex AI (LLM inference) — under a zero-retention agreement: your prompts and completions are not stored at all, including for abuse monitoring. Google approved our exception to its prompt-logging policy on September 15, 2026; before that date content could be retained for a limited period for abuse detection, and never for training.
• Amazon Web Services, through Amazon Bedrock (LLM inference for the Claude models, and for Gemma when Google Cloud cannot serve a request) — your prompts and completions are not stored or logged, and the model’s developer has no access to them. AWS’s automated abuse detection keeps one exception: apparent child sexual abuse material in an image may be stored, reviewed and reported to the authorities. Claude requests may be processed in any supported commercial AWS Region.
• Prompt caching. Where a model supports it, the provider may keep the beginning of a prompt for a short time, scoped to your account, so a request repeating it costs less: in volatile memory only on Google Cloud, and for five minutes after its last use on the Claude models. You can turn it off for any request with "prompt_cache": false.
• We no longer offer OpenAI models: they were removed on November 3, 2026, because OpenAI’s terms did not provide zero retention. No request reaches OpenAI.
• Screening, and where it applies. Only requests to a provider without a zero-retention basis may be classified before they are forwarded. None is used today, so no request is classified. Requests to any zero-retention model are never classified, and we will not extend screening to them: inspecting that content would contradict the agreement that makes those models zero-retention in the first place.
• If a provider reports misuse. We can tell them which API key was involved, how much traffic it sent and when, and we can revoke that key immediately. We cannot give them the requests themselves, because we do not have them. We may also suspend a key automatically if it crosses an abuse threshold; that is reviewed by a person and can be reversed, and we will tell you it happened.
• The operator that served your call is named on each receipt, and Schedule 3 of our DPA states the current basis for each one.
• Brevo (transactional email)
• The hosting providers that run our backend and this site
• Legal authorities, if required by law

6. Data Retention

We retain your account and project data while your account is active. After account deletion, residual logs may persist for up to 30 days for operational reasons, after which they are removed.

The inference request metadata described in section 2(d), and any record of an automated suspension, are kept for twelve months and then deleted. That length is set by what it is for: a model provider investigating misuse may reach us well after the requests in question. Your prompts and completions are not part of these records, and our model providers do not retain them. The only content a provider may keep is apparent child sexual abuse material in an image sent through Amazon Bedrock (the Claude models, and Gemma when Google Cloud cannot serve a request), which AWS may store, review and report to the authorities (section 5). Prompt caching, also in section 5, is not a record: the beginning of a prompt is held briefly so a repeated request costs less, and then dropped.

offrec chat conversations, files and projects are kept, encrypted, until you delete them or your account (section 2(f)). Deleting removes them from our database at once; encrypted copies can remain in our database provider’s backups until those backups expire.

7. Security

We implement reasonable safeguards to protect your data, but no system is completely secure.

8. Your Rights & Choices

Depending on your jurisdiction, you may request to access, correct, or delete your data. You can also:

• Permanently delete your account, all projects, and all uploaded data from your account settings (this is irreversible)
• Opt out of product analytics and session recording at any time by enabling your browser’s “Do Not Track” or Global Privacy Control setting—we honor it automatically and stop collecting analytics for that browser

9. Cookies / Tracking

We use cookies and similar technologies for:

• Authentication and keeping you signed in
• Product analytics (PostHog), subject to the opt-out above
• Website analytics and ad attribution (Vibiz) on offrec.ai, subject to the opt-out above

10. Contact

Questions about privacy can be sent to [email protected].