Air-Gapped Deployment & Uncensored Local LLMs

For many enterprise and government clients, sending sensitive corporate data to external cloud LLM providers (like OpenAI or Google) is a non-starter due to strict compliance, privacy, and security regulations.

To address this, Jorvis supports full Air-Gapped Deployment.

100% On-Premise Operation

The entire Jorvis stack—including the database (PostgreSQL + pgvector), the application backend, the GraphRAG engine, and the Large Language Models—can be deployed within a customer's private network or Virtual Private Cloud (VPC), completely isolated from the public internet.

The Abliterated Qwen 3.5 Advantage

A common challenge with deploying open-weights models (like Llama 3 or standard Qwen) in an enterprise setting is "moral refusals" or "alignment tax." These models are heavily aligned to refuse queries that sound vaguely sensitive, even when the query is completely benign in a business context (e.g., "Analyze the kill chain of our cybersecurity events" or "Show me the termination list for this quarter").

Jorvis utilizes an Abliterated (Uncensored) Qwen 3.5 Local-LLM. "Abliteration" is a technique that removes the model's refusal mechanisms without degrading its underlying reasoning and coding capabilities.

This guarantees that the model will always attempt to answer queries based on the provided corporate data, completely eliminating frustrating "I cannot assist with that" responses during critical data analysis tasks.

Security by Isolation

Because the model is air-gapped and uncensored, the security model shifts from "hope the LLM doesn't leak or refuse" to "strict network and access control." Jorvis enforces security at the boundary (RBAC, Row-Level Security, and SQL Validation Guards), allowing the LLM to freely operate as a pure reasoning engine within its sandboxed environment.