Hermes Investigative AI

Answers analysts can defend.

Hermes is an investigative AI application for analysts, targeters, and operators. It answers from your own material with citations, says “Insufficient evidence” when there is nothing to cite, gates analytic assertions with tradecraft and OPSEC policy, and installs on a single Linux host with no internet connection.

Request a Briefing Built on ARIA: policy, approvals, audit, and replay

What analysts do

From raw files to a reviewed output.

Step 01

Bring your material

Create a project and add the documents, spreadsheets, and logs your team already works from.

Step 02

Investigate and develop findings

Ask for cited answers, extract entities and relationships, each showing the passage it came from, and record assertions you can review, challenge, and revise. Turn the work into repeatable workflows that run on demand or on a schedule.

Step 03

Review and publish

Outputs move through human review and approval before publication.

Built-in controls

Evidence first, policy enforced, people in charge.

Evidence

Answers cite your files, or say “Insufficient evidence” instead of guessing. An assertion needs at least one evidence record. Citations show where a claim came from, not that it is correct: language models can still be wrong, so people review the work.

Policy and approval

Tradecraft and OPSEC policies can warn, require approval, or block, including on model calls that carry project content. A run needing approval pauses until a person decides; if policy cannot be evaluated, the default is to block. Hermes advises; people approve, publish, and decide. Policy and approvals

Explore evidence and analysis

Configured patterns

Two shapes of work it is built for.

Pattern 01 · Adjudication

Reviewing a package against a standard

A workflow checks each unit of a package against the framework text for a present narrative and sufficient, specific, traceable evidence. Findings become evidence-linked assertions routed to human review.

Pattern 02 · Multi-agent investigation

A team of agents over a technical corpus

Analysis, execution, and research agents correlate tool output, run checked commands in isolated environments, and curate cited write-ups. Any step that expands scope waits for a scoped human approval.

Product capabilities described as patterns, not case studies of a named customer or program.

Where Hermes runs

Your infrastructure, your models.

Hermes installs on a single Linux host from a package designed for disconnected networks. With self-hosted models and offline mode turned on, it refuses connections to non-local hosts; isolate the network as well. Bring your own self-hosted models.

Explore edge and offline deployment

Briefing

Bring a case you already have.

The most useful briefing is one where Hermes works a problem your team already handles manually. Hermes runs on ARIA; the Hermes white paper has extended detail.

Request a Briefing