Built for EU AI Act obligations

Intelligence that can be governed

VIRAM checks every claim the model makes about your documents against your documents — and tells you, in the answer, where it got something wrong. You get the analysis and a list of what could not be verified. Nothing is quietly dropped. Nothing is quietly wrong.

Ingestion Verified Against Source
21,274
Table cells, all extracted exactly
0
Words missing across 3,142 pages
1–2ms
Source lookup, any file
~$0.03
Per query, 1.38M-word corpus
🔒
Air-gap deployable
Nothing hidden
~90%
Processing in Code, Not Model Inference
3,142
Pages Verified Against the Source PDFs
Zero
Model Calls to Retrieve a Stored Fact
Yours
Model, Hardware and Jurisdiction

Ingest once, query forever

Most AI systems re-read your documents on every question and forget everything in between. VIRAM reads them once.

📁

One-Time Ingestion

Files are ingested once. File name, structure, section, paragraph and page are written onto every chunk at that moment and never recomputed. No re-processing, no re-embedding, no recurring ingestion cost — the corpus is prepared, permanently.

🔗

Cross-Document Queries

Ask one question across your entire corpus. There is no fixed document limit: work is batched, so the constraint is how long you are willing to wait, not the model's context window. Find contradictions, trace themes, compare positions across everything at once.

🧠

Memory the Model Doesn't Hold

Your documents live in a separate store, not in the model's context and not in its training. The model never has to remember anything. Attach files or query from memory — the answer is built from the same content either way, and the store outlives any model you choose to run.

Built for inquiry, not apology

When a minister asks how a conclusion was reached, when an FOI request demands the reasoning trail, when a Royal Commission examines the evidence — VIRAM provides answers, not excuses.

🔍

Full Traceability

Every conclusion carries the file, the structural location and the page it came from. Every prompt and response is logged with a timestamp. Any cited passage can be retrieved and read in 1–2ms. Entries are chained with SHA-256 — each one hashes the one before it, so a modified record breaks the chain.

Verified Outputs

Every claim attributed to a source document is checked against that document before you see it. Quotes are matched against the text. Cited locations must exist and must contain what the claim says they contain. Where a wrong citation can be corrected, it is. Where it cannot, the system says so and names the place in the source for you to check.

⚖️

It Tells You When the Model Was Wrong

The verification layer sits above the model, not inside it — so it reports on whichever model you run. When output is softened, misattributed or invented, the notice names the failure and identifies the model as the cause. You are not asked to trust the model. You are shown where it fell short.

Intelligence governed outside the model

Around 90% of the work happens before the model is invoked — in deterministic code you can inspect, log and repeat. The proportion varies by query.

1

Deterministic Pre-Processing

Documents are ingested once, stored permanently and retrieved without involving a model. Extraction, structural labelling, chunking, retrieval and verification are code, not inference. Retrieving a stored fact costs nothing and returns the same result every time.

2

BRAIN Reasoning Layer

BRAIN assembles what the model receives: structure, location and content drawn from across the whole corpus. Where content exceeds what a single call can carry, it batches and synthesises across batches rather than sampling. It reads every part of every document — not the opening, a little of the middle and the end, which is what a long context window does in practice.

3

Constrained Model Invocation

The model reasons from retrieved source content, not from training memory, and receives pre-structured material rather than raw text. Models are interchangeable: run an open-source model on your own hardware, or your organisation's approved API. Nothing in the architecture depends on which one you choose.

4

Post-Inference Verification, and an Honest Limit

Output is verified against source before delivery, with up to three correction cycles. What cannot be verified is listed for you, not removed. And if answering your question properly would exceed the model's context window, VIRAM does not return a partial answer — it states the size of the request against the limit, and stops. A confident half-answer is the most dangerous output an AI system can produce.

Two capabilities, one governable architecture

Document intelligence over your own corpus, and research that brings vetted external science into it.

Core

VIRAM Core

Governable document intelligence

  • Ingest once, query for as long as you keep the files
  • 1–2ms retrieval of any ingested content, checkable by you against the answer
  • Every citation carries file, structure and page
  • SHA-256 chained records — tampering breaks the chain
  • No model call, and no token cost, to retrieve stored content
  • Runs on modest hardware; deployable fully air-gapped with a self-hosted model
Request Information
Research

VIRAM IIQ

Infinite IQ — research acquisition and synthesis · in development

  • Everything in Core, plus:
  • Papers retrieved from 20–30 vetted scientific sources
  • Papers only — never the open web, never general web pages
  • New material joins your existing corpus and is analysed alongside it
  • Branching research lines with continuation prompts
  • Requires network access to those sources, and a model with strong scientific knowledge — unlike Core, it is not intended for air-gapped deployment
Request Information

Your infrastructure, your jurisdiction

Deploy on your servers, run the model you have approved, keep your data inside your borders. Core is designed for organisations that cannot compromise on sovereignty or auditability.

🔒
Air-Gap Deployable (Core)
Operates fully offline with a self-hosted open source model
🔌
Model Agnostic
Bring your own model — the verification layer is independent of it
📋
Full Log Transparency
Every operation logged, and the logs can be made available to you
🗑️
User Data Control
Delete anything, gone permanently. No silent retention.

Built for the obligations, not the badge

Transparency (Art. 13)
Human-readable reasoning trails, and an explicit list of what could not be verified
Record-keeping (Art. 12)
Every prompt, response and operation logged, chained and inspectable
Human Oversight (Art. 14)
Outputs are checkable against source in 1–2ms; correction needs no retraining or redeployment
Data Governance
Your documents remain the only source of truth; the system does not browse the open web

Ready to look at VIRAM properly?

VIRAM is in active development and we are talking to organisations interested in early deployment. If your decisions have to withstand scrutiny, we would welcome the conversation.

📍
London
United Kingdom
📍
Australia
Brisbane

Request a Private Briefing