Security approach
Where your data goes, decided up front.
Privacy in an AI system is a set of choices at each stage, not a single switch. This page explains how we scope those choices with you.
Deployment options
Three places a system can run.
A server you own
The system runs in an environment under your control. You decide who has access and where the data lives. Capacity, such as GPU hardware for larger models, is checked during scoping.
A dedicated environment
The system runs in an environment set up for you alone, operated under an agreed arrangement. Responsibilities for access, updates and backups are written down.
An approved API provider
A hosted model is used under terms you have reviewed. We help you read what the provider says about retention and training, and design the system so sensitive material is not sent unless you have approved it.
Which one fits depends on the sensitivity of the material, your obligations and what you operate today. We recommend the simplest option that meets the requirement.
The data boundary
Six stages that touch your data.
For each project we write down, for every stage, where it runs, who can see the data and how long it is kept.
- 01
Extraction
Reading text out of files, scans and web pages. It may run in a tool or service separate from the model.
- 02
Embeddings
Turning passages into numeric vectors for search. This may use a different model, in a different place, from the answering model.
- 03
Retrieval
Looking up passages at question time. Access rules are applied here, so a person only retrieves what they may read.
- 04
Inference
The model that writes the answer. Where it runs, and what the provider keeps, is the question most people ask first.
- 05
Logs
Records of questions and answers. They can contain the same sensitive material as the documents, so they need a retention rule.
- 06
Backups
Copies of the data and the index. They live as long as the policy says, which may be longer than you expect.
How we build
Controls we design in.
- Access follows the permissions you already have. A person who cannot open a document cannot get it through the assistant.
- Where a wrong action is costly or cannot be undone, a person approves it first, and sees the evidence they need to decide.
- Answers point to the passages they used, and the system can say that it did not find an answer.
- Actions that must not run twice are protected against duplicates and retries.
- Secrets, keys and client documents stay out of source control, screenshots and examples.
What we do not claim
Honest limits.
- We do not claim certifications we do not hold, and none are stated on this site.
- We do not promise perfect accuracy, zero hallucinations or universal privacy.
- We do not claim that running a model on your own server settles the whole privacy question.
- We do not process confidential documents in a public demonstration or a website form. Use synthetic or approved sample data when exploring an idea.
One question
Which manual process currently takes the most time from your team?
Answer in a sentence or two. An engineer replies with where we would start, what it takes and what we would not automate.