Privacy Policy
Spanda Works LLC · Version 3.1 · Last updated November 14, 2026 · Takes effect for your account when you accept it · [email protected]
Until you accept it, version 2 (October 29, 2026) applies to your account.
This policy explains what Tīrtha collects, where it goes, how long we keep it, and what you can do about it. It replaces the July 4, 2026 policy. Spanda Works LLC, 6650 Rivers Ave STE 100, Charleston, SC 29406, USA, is responsible for your data (the "controller" under EU and UK law).
This page describes the system as it runs today. Where we are building something stricter, section 15 says so, and this page changes only when that is live.
The short version
- Your request goes through our gateway to a company that runs the AI model (a "host"), and the answer comes back through us.
- Our request log holds numbers, not content: time, tokens, cost, which level or model served it, whether it was checked.
- When an answer passes our check, we may store the prompt and the answer in a shared cache, and serve that answer to other customers. You can turn this off for any request (section 3).
- Tīrtha takes text, images, audio and video (section 2).
- We do not use your prompts, code, images, or output to train or fine-tune models.
- We do not sell your data and we do not use it for advertising.
1. What we collect
| What | Why | Kept for |
|---|---|---|
| Account: email, name, Google sign-in identity | run your account, send notices | while your account is open, then 90 days |
| API keys, stored as a hash. A key made by connecting a tool is held in plain form only until the tool collects it, or the request expires | check that a request is yours | until you revoke the key or close the account |
| Terms acceptance record: which version, when, the IP address, browser and country at that moment | prove what you agreed to; sanctions rules | as long as we may need to prove you accepted, including after your account is deleted (at least the life of the account plus 7 years) |
| Payments: amount, date, fees, tax, card country, Stripe reference. Stripe handles the card or wallet details; we never see a full card number | billing, refunds, tax, fraud prevention | 7 years (tax and accounting law) |
| Request log, every request: time, a hash of your key, level or model, host, tokens, cost, speed, whether it was checked, request id. No prompt or answer text. | billing, limits, abuse prevention, quality | no automatic deletion yet; we delete on request where we can (see section 15) |
| Shared cache: the prompt and the answer, when the answer passed a check and you did not turn the cache off | answer repeat requests faster and for less | see section 3 |
| IP address and basic network data. Our gateway's request logs do not record client IP addresses; IP addresses can appear in our servers' system logs and in Cloudflare's logs | security, rate limits, blocking sanctioned locations | our system logs are removed as they rotate, not yet on a fixed schedule (see section 15) |
| Support email | answer you | 2 years after the conversation ends |
| Website analytics: page views and referrers, with no cookies (Cloudflare Web Analytics); a campaign tag kept in your browser's session storage | learn which pages people read | Cloudflare's standard period |
Cookies. We use only the cookies needed to keep you signed in and keep the site secure. Our analytics use no cookies. We do not track you across other sites.
2. What happens to a request
- Your request reaches Cloudflare, which runs our network edge. Cloudflare decrypts it there and forwards it, encrypted, to our gateway.
- Our gateway runs on Amazon Web Services in the United States.
- We send the request to a host to run the model, and sometimes to a second model to check the answer, or to a higher level if the check fails. One request can reach more than one host. To find matching cache entries we may turn the prompt into an embedding using Amazon Bedrock.
- We return the answer with a label and the charge, write one log row of numbers, and may cache the answer (section 3).
Images, audio and video. Once you accept our Terms, including the Media Terms, you can send images, audio or video and receive generated images, audio or video. These files are handled like prompts and answers: they go to the host that runs the model, and section 6 says how long hosts keep data. We keep no copy of the files you send or the files a model makes; our job record keeps only what billing needs (the model, the time, the size and the charge). A generated file comes back in the response or as a link from the host, and the link may stop working after 24 hours or sooner. Speech is turned into text only: we do not create voiceprints, and we do not use images, audio or video to identify anyone.
3. The shared cache
- When an answer passes our check, we may store it with the prompt that produced it, because the prompt is how the entry is found again. The prompt is not redacted. If it contains code, file paths, or business logic, that text is stored.
- The cache is shared by default. Another customer who sends the same or a very similar prompt may receive the cached answer. Your identity is never served with it.
- Turning it off. Send
tirtha.no_cachein the request body. When it is set, nothing is read from the cache and nothing is written to it for that request. Private cache scopes are available by arrangement. - How long. Today, entries have no automatic expiry: they stay until we delete them. A 90-day limit is coming (section 15).
- Deleting yours sooner. Email [email protected]. We delete the entries we can trace to your account within 30 days. Older entries, made before we began tagging entries by account, cannot be found by account; we can delete them if you tell us the prompt text. Answers already served to others before deletion cannot be recalled.
4. Abuse checks
We and our hosts may run automated checks on requests and output to find misuse, such as the uses our Terms forbid. If a check flags a request, a person at Spanda Works may review what we hold about it: its log row of numbers and, if its answer was cached, the cached prompt and answer. We do not keep a separate copy of flagged requests. Any notes we make about one are kept no longer than 90 days unless the law requires longer. If we find apparent child sexual abuse material, we report it to the National Center for Missing and Exploited Children.
5. Training
We do not use your prompts, code, images, or output to train or fine-tune models. Cache entries made from customer traffic are marked as never usable for training, and our training-export tool refuses them.
The hosts we use agree in their terms not to train on our API traffic, with two exceptions that have not yet confirmed this in writing: one host that serves some Max-level requests, and one host for models outside its zero-retention option. We have asked both, and we will update this section when they answer or when we stop sending them customer traffic.
6. Who processes your data
These companies process data for us. Some hosts' contracts do not allow us to name them publicly, so hosts are listed by type and region. Business customers can get the named list under a confidentiality agreement by emailing us.
| Who | What they do | What they see | Where |
|---|---|---|---|
| Model hosts | run the model that answers or checks a request | prompt and answer text | United States, European Union, Singapore; one host may process requests in other regions outside mainland China, and stores data in Singapore |
| Amazon Web Services | servers, database, embeddings (Bedrock), some model steps, email | everything we store; prompts for embeddings and model steps | United States |
| Cloudflare | network edge, tunnel to our servers, website hosting, analytics, blocking | requests as they pass through, page views | worldwide edge network |
| Stripe | payments, refunds, receipts | payment and billing details | United States and where Stripe operates |
| sign-in | your email and name, at sign-in | United States |
Backups and other companies' logs. Deleted data can remain in our encrypted database backups until those backups expire on their normal schedule. Cloudflare, AWS, and the model hosts keep their own logs (for example IP addresses and request details) under their own retention rules, which we do not control.
How long hosts keep data. Most hosts we use keep no prompts after a request ends. Some keep request logs for a short time for security and debugging under their own terms; Amazon Bedrock may keep some requests up to 30 days to detect abuse. One model host keeps prompts and answers for up to 30 days, and keeps content its safety filter flags for up to 2 years.
Changes. We list each new sub-processor here before it handles customer content. If it works within the regions and time limits this page already describes, or handles only requests you choose to send to a model or feature that says so, the listing is our notice. Otherwise it is a material change under section 14. If you object to a change, you may close your account and we refund your unused paid credits.
7. Where data is processed
Our servers and database are in the United States. Depending on the model, a request may be processed by a host in the United States, the European Union, or Singapore, and one host may process requests in other regions outside mainland China. We do not offer a choice of region today. Wherever you are, your data may be transferred to and processed in these regions. Where the law requires a safeguard for these transfers, we rely on the standard contractual clauses in our providers' data processing terms.
8. Why we are allowed to use it (EU and UK)
- To perform our contract with you: running your account, serving requests, billing, support.
- Legitimate interests: security, abuse prevention, improving the service, the shared cache (which you can opt out of per request), usage numbers combined across customers.
- Legal obligation: tax records, sanctions rules, responding to lawful requests.
- Consent: anything optional we ask you to agree to separately.
9. Your choices and rights
- Keep a request out of the cache: send
tirtha.no_cachewith it. - Access, correction, deletion, export, objection: email [email protected]. We answer within 30 days, and tell you if we need longer and why. We may ask you to confirm the request comes from your account email.
- EU and UK: you may also restrict processing and complain to your data protection authority.
- California and other US states: you can ask what we hold about you and ask us to delete or correct it. We do not sell or share personal information for cross-context advertising, and we will not treat you differently for using these rights. We treat Global Privacy Control and Do Not Track signals as an opt-out where the law asks.
- Data processing agreement: business customers who need one under GDPR or a similar law can email us to arrange one.
10. Usage numbers we publish
We may publish statistics about Tīrtha, such as speed, cost, and how often checks pass. We publish them only combined across many customers, in a form that does not identify you and contains none of your content.
11. Requests from governments
We give customer data to a government or court only when the law requires it. Unless the law or the order forbids it, we tell you first so you can object.
12. Security
Data is encrypted in transit and at rest. API keys are stored as hashes. Access to production runs through AWS access controls, and our servers have no open SSH ports. No system is perfectly secure. If we learn of a breach that affects your personal data, we will tell you without undue delay and within any deadline the law sets, with what we know and what we are doing about it.
13. Children
Tīrtha is not for anyone under 18, and we do not knowingly collect their data. If you think a child has given us data, email us and we will delete it.
14. Changes to this policy
We add models, kinds of models (such as image, audio or video) and hosts often. These are product updates. They take effect when we post them here, as long as either (a) the new host processes data in a region this policy already lists and keeps it no longer than this policy already says, or (b) the change only affects requests you choose to send to a new model or feature, and that model or feature says so before you first use it. A change that reduces the protections for data you already send us, such as keeping it longer, sending it to a new region, or allowing a new use of it, is a material change. For a material change we will email you at least 14 days before it takes effect, or ask you to accept it before your account can make more requests. A change that only adds protections takes effect when we post it. Each version shows its date.
15. What we are building next
These are built but not yet running. Until this page says otherwise, sections 1 to 3 above describe what happens today.
- Cache limit. Entries made from customer traffic stop being served, and are deleted, 90 days after they were last written. Older entries made from customer traffic are deleted when this starts.
- Log limits. Request log rows deleted after 12 months, and our servers' system logs, where IP addresses can appear, kept no longer than 30 days.
- More ways to turn the cache off. A request header,
X-Tirtha-Cache: off, and a setting that turns the cache off for a whole API key. - Images. Image input for models that can read images. Images will be sent only to the hosts that process the request, never put in the shared cache, and never written to our logs (the logs will record only that an image was present, its type and its size).
Contact
Spanda Works LLC, 6650 Rivers Ave STE 100, Charleston, SC 29406, USA · [email protected]