We go to considerable lengths to keep user data safe.

Kelsen has partnered with Private AI, to deploy their cutting-edge machine learning identification technology. This model relies on the latest advancements in machine learning to identify personally-identifiable information based on context, understanding text in a way that is similar to humans. This technology identifies over 55 different types of direct and quasi-identifiers in 49 different languages.

Importantly, this identification model is hosted on Kelsen’s server infrastructure, which means that the data never gets transferred to Private AI or any third parties.

To ensure informational security, all written data submitted to or produced in Kelsen software:

(i) is verified for over 55 different types of personally-identifiable information;

(ii) any personally-identifiable information found in the input data (such as names, dates, document numbers, addresses, countries, email, IP, monetary values, etc.) is automatically extracted and substituted for context-appropriate pseudonyms;

(iii) the correspondence keys linking the pseudonyms to the original information are kept on-premise, in an encrypted repository that is only accessible to the user and these data are never transferred to a language model; and

(iv) the pseudo-anonymized data is only re-identified with the original personally identifiable information in the last leg of the text generation process, after the output text has been generated by the language model, and before the data is served back to the user.

Presently, Kelsen employs proprietary models comprising neural networks and exclusive custom builds of open-source Llama 3.2  large language models, hosted on Kelsen’s cloud infrastructure.

No. We do not use client data to fine-tune or train our base model, although clients are free to safely finetune their dedicated instances of our models with their own data.

Client fine-tuning of their respective instances is strictly limited to the client’s own instance and does not communicate with is visible to, or can otherwise benefit Kelsen’s base models or the model instances of any other client or third party.

Moreover, user’s prompts (inputs) and completions (outputs), embeddings and training data:

(i) are NOT available to other customers;

(ii) are NOT available to Kelsen or any third parties;

(iii) are NOT used to improve language models deployed by Kelsen or any third parties;

(iv) are NOT used to improve products or services provided by Kelsen or third parties;

No. We do not use client data to fine tune or train any instance of any language model, including our own language models and/or commercial models like GPT-4 and others.

Anonymized and sanitized user data within specific user-managed Projects are only used to train Project-specific instances of our proprietary Logical Thought Network (LTN), a neural network, which does not communicate with those of any other Projects. When a Project is deleted, so is that Project’s instance of the LTN, along with all submitted data.

User’s prompts (inputs) and completions (outputs), embeddings and training data:

(i) are NOT available to other customers;

(ii) are NOT available to Kelsen or any third parties;

(iii) are NOT used to improve language models deployed by Kelsen or any third parties;

(iv) are NOT used to improve products or services provided by Kelsen or third parties;

All data is encrypted in transit and at rest in our system.

Kelsen encrypts all data in transit via SSL/TLS (Transport Layer Security) with the 256-bit AES (Advanced Encryption Standard) protocol, meeting industry standards for externally-facing systems.

AES 256-bit encryption with a rotational KMS (Key Management Service) is used to protect data at rest.

Encryption keys, including the right to revoke said keys, may be assigned to the company-designated administrators.

keys will be stored in an encrypted backup vault.

Kelsen Draft is a cloud-based application.

User data is hosted and stored in Kelsen’s cloud infrastructure, in servers located in or near to the user’s physical location, in compliance with national data residency regulations, such as GDPR, PIPEDA, LGPD, etc.

For high-volume clients, cloud infrastructure and hosting may be changed to a specific cloud service or the client’s own cloud infrastructure, on a case-by-case basis, to accommodate specific requirements and data management policies.