Security

Security you can audit, not adjectives

Sovereignty first — where your data lives, who can reach it, and the evidence your review team actually checks behind our AI and RAG deployments.

AI illustration

This page is written for the reviewer, not the browser. Security starts with a simple question: where does your data live, and who can touch it. That answer comes first, then every control a regulated buyer's security assessment, vendor risk questionnaire, or EU AI Act / DORA / GDPR mapping actually checks. Each claim below maps to a mechanism we can demonstrate to your auditor — a database policy, a redaction step, an audit log that cannot be edited, a deployment topology — and nothing more. We host in the EU, and with an EU model provider or your own server, your documents and prompts stay in the EU while the AI works. Each claim is labelled for what it is: enforced, a property of the design, or on the roadmap and not yet held. This page is engineering, not legal advice — a description of how we help you meet and prepare for the EU AI Act, DORA, and GDPR, grounded in what we deliver.

Features

  • Data stays in the EU, inside the boundary you set

    You choose the deployment boundary, and the AI works inside it. Your documents, index and audit trail are hosted in the EU, and model execution stays in the EU when you choose an EU provider, the EU regions of your cloud or your own model. EinCoreRAG Enterprise is installed on your own server, inside your network — with our team.

  • Database-enforced tenant isolation, described precisely

    Isolation is pushed down into the database, not left to application code. The database filters every query by customer, tested on 1,000+ workspaces, and Enterprise adds a separate search index per customer, so a misconfigured query does not reach another customer's data. This is a fail-closed model, enforced in the data engine rather than promised in prose, and we document exactly which model your deployment is on.

  • PII redacted before the model sees your data

    Sensitive fields are redacted at the input boundary, before the model ever processes a request, so personal data does not reach the inference layer in the clear. This keeps identifiable information out of inference and out of the context window, advancing your data-minimisation posture and narrowing what personal data GDPR ever puts in scope at the boundary rather than trusting that to a prompt. We can show your reviewer the redaction step in the request path and what it removes.

  • Prompt-injection detection and denial-of-wallet caps

    Every input and output is screened for prompt injection — instructions that try to override the system's behaviour or pull out context — and what is caught is contained before it reaches the model. Every write action into your business systems still waits for a person to approve it. Per-tenant spend and rate caps bound token consumption, so a malicious or runaway workload cannot quietly burn your budget or your inference capacity. Both are controls you can inspect, not policy statements.

  • Tamper-evident audit trail that cannot be edited, enforced by the database

    Every request, retrieval, and answer is written to a tamper-evident, per-tenant audit trail whose entries the database itself refuses to change or delete, kept for a fixed period. It records who asked what, what context was retrieved, and what the system returned — the record your assessor asks for. The evidence is structured so it can be mapped to SOC 2, ISO 27001, and EU AI Act reviews, including the logging and traceability the AI Act expects of deployers of high-risk systems under Article 26.

  • Grounded, closed-context RAG with human handoff

    Answers are generated only from your retrieved documents in a closed context, with citations back to source — not from open-ended model recall — which narrows hallucination and keeps every response traceable. When confidence is low or the question falls outside the grounded context, the system hands off to a human rather than inventing an answer. That boundary is configurable and logged, so human oversight is a control you can evidence, not an assumption — which is what oversight obligations actually require.

  • Certifications and attestations

    Our controls are mapped to SOC 2 and ISO 27001 evidence, and attestation is on our roadmap — it is not yet held, and we will not claim a certification we do not hold. If your questionnaire asks for a certificate we cannot produce, you get a straight answer and the control-to-evidence map that stands behind it.

  • Regulation framed as engineering — with evidence packaged for your assessment

    We help you meet and prepare for the EU AI Act (Regulation (EU) 2024/1689), DORA (Regulation (EU) 2022/2554), and GDPR through concrete delivery: risk-tier and provider-vs-deployer mapping, logging and human-oversight hooks, ROPA and DSAR artefacts, and reduced critical-provider and concentration risk. The deliverable is a review pack, not a demo — generated evidence for a real use case. This is engineering support for your compliance work alongside your own counsel, not legal advice or a guarantee of compliance.

FAQ

  • Where does our data actually live, and does any of it leave the EU?

    Your documents, index and audit trail are hosted in the EU. Model execution stays in the EU when you choose an EU provider, the EU regions of your cloud or your own model. EinCoreRAG Enterprise is installed on your own server, inside your network — with our team.

  • Do you hold SOC 2 or ISO 27001 certification?

    Our controls are mapped to SOC 2 and ISO 27001 evidence, and we will walk your assessor through that mapping control by control. Formal attestation is on the roadmap and not yet held. What is available to you today is the control-to-evidence mapping itself and a tamper-evident per-tenant audit trail structured for those frameworks.

  • Can you guarantee one tenant can never reach another tenant's data?

    We do not use the word guarantee for absolutes. Multi-tenant deployments enforce isolation in the database itself, tested on 1,000+ workspaces, and Enterprise gives each customer its own search index, producing a fail-closed model where the database filters cross-tenant records at the query boundary. We do not, however, market this as an absolute guarantee. Where you need separation that is physical rather than only enforced in software, EinCoreRAG Enterprise is installed on your own server, inside your network — with our team — and we document precisely which model your deployment uses.

  • How does this help us meet the EU AI Act, DORA, and GDPR?

    Through engineering, not legal advice. We help you map use cases to AI Act risk tiers and provider-vs-deployer duties, produce the logging and human-oversight the Act expects of deployers (including Article 26 for high-risk systems), generate ROPA and DSAR artefacts and enforce data residency and redaction for GDPR, and reduce the critical-provider and concentration risk DORA scrutinises for financial entities. This is how we prepare your deployment to meet the requirements alongside your own counsel; it is not legal advice and not a compliance guarantee.

  • What does our security review get to see before we commit?

    A working deployment on your own documents, and the evidence to check it line by line. We set up one real use case — in the EU cloud, or on your own server, inside your network — with the success metric agreed up front, an AI Act deployer evidence pack (Art. 26 for high-risk use) and access to the audit trail, so your review team evaluates real evidence, not slideware, before there is any wider commitment.

Bring us the project your compliance team said no to

The fastest way to clear a security review is to run the system where your data stays. Start with one real use case on your own documents — in the EU cloud, or on your own server, inside your network — and your review team gets the evidence pack and the audit trail: proof your assessment can check line by line, not slideware. Request a call, or write to us first at teamtechrevati.com.

Request a call