How CogniGrid keeps data inside
Every request passes through two systems you host yourself: a Neural WAF that inspects traffic, and a Private RAG pipeline that reads your documents.
Layer-1 Neural WAF: request inspection
-
Ingress capture
Prompts and API calls are intercepted at the network edge, before they reach any model.
-
Neural threat scoring
A classifier scores each request for prompt injection, jailbreaks and attempts to extract sensitive data.
-
Policy decision
Safe requests pass through. Risky ones are blocked or redacted, and every decision is written to an audit log.
-
Local model inference
Approved requests reach an LLM running on your own GPUs. Responses are scanned again on the way out.
Private RAG: from PDF to answer
Parse
PDFs are read and chunked in memory on your server.
Embed
A local embedding model turns text into vectors. No API call.
Store
Vectors sit in an encrypted store inside your perimeter.
Answer
The local LLM answers with citations to your source pages.
CogniGrid vs. public AI APIs
| Capability | CogniGrid AI | Public APIs |
|---|---|---|
| Data leaves your network | Never | On every request |
| Works offline (air-gapped) | Yes | No |
| Prompt injection defense | Built-in Neural WAF | Basic or none |
| Document retrieval | On-prem Private RAG | Files uploaded to vendor |
| Audit logs | Full, exportable to your SIEM | Limited |
| Data residency control | You choose the site | Vendor decides |