Neural ICE
EN · FR · DE · IT Talk to us
Secured across the device lifecycle · technical brief

Ready for your
security review.

For security and IT leaders: what stays on site, what leaves and why, how the platform is protected, and what ships with every Neural ICE appliance. Open-weight model of your choice, on hardware you host.

Local inference · air-gap deployment as an option ICE-CoreOS is open-core Designed for GDPR · Swiss FADP · AI Act · NIS2 environments
Data boundaries

Your knowledge stays on the device.

Documents, knowledge, prompts, answers and inference stay on the appliance, including air-gapped. Local mode: prompts and answers only to your authenticated app.

Control · licence & releases
The appliance contacts Neural ICE only for licence checks and signed updates (licence data, hardware fingerprint, device name, platform type). No documents, chats or usage telemetry. Updates arrive signed; air-gapped sites use signed offline media.
Nomad Access · optional
Off by default. Paired devices reach the appliance over an encrypted private network (direct when possible, else Swiss or EU relay). Each device must authenticate; nothing on the public Internet. Off in air gap.
Events to your SIEM · optional
You can send security events to your SIEM, or keep a tamper-evident audit journal your security team can read. That is your choice, not customer content leaving on its own. It does not pass through Neural ICE.
Connectors · optional
Google Workspace or Microsoft 365 only when you enable them, under your credentials, for the fields needed for that action. The full knowledge base never leaves. Off by default, revocable anytime.

Anything not listed here is not part of the product.

Countermeasures · platform

Secure by design, before the first prompt.

Machine, storage, access and communications first; AI controls next. Updates are signed and verified before install.

Verified boot chain Immutable, signed OS · retained local rollback Hardware-bound disk encryption Mandatory access control Encrypted traffic between services No service intended for public Internet exposure Multi-user on one appliance · capacity set by the machine Tamper-evident audit trail Air-gap capable Send events to your SIEM
Neural ICE · on your premises

What you get on the appliance.

Three formats · delivered preinstalled by Neural ICE ICE-CoreOS is open-core · inspect it on GitHub ↗

Three formats, one product

Desk on NVIDIA DGX Spark · Workstation on NVIDIA RTX PRO 6000 · Station on NVIDIA DGX Station (coming soon) · team size sets the format.

why it matters

No datacenter required. We choose, prepare and test the hardware; you choose the size.

ICE-CoreOS

Signed, immutable OS with local rollback · open-core, inspectable on GitHub.

why it matters

Verified boot, atomic updates, previous deploy kept for local recovery.

Open-weight models you choose

Smaller or larger open models, validated on the device.

why it matters

You pick the model. The active version stays identifiable, local and replaceable.

Sealed on your premises

Hardware-protected storage · no public Internet exposure.

why it matters

Keys stay on the device. Authenticated local access. Tamper-evident audit trail.

From documents to deliverables

Your documents and data become source-grounded analyses and editable files.

why it matters

Agents divide work; tools calculate; sources stay inspectable.

Fits what you already use

Native client, Nomad Access for remote without public exposure, modular connectors (mail, calendar, Microsoft 365, MCP) when enabled.

why it matters

Local by default. Remote without public exposure. Connectors only when you turn them on.

Takes your work with you

Documents, archives, configs and adapted models leave with you.

why it matters

You own the hardware; the OS core is open; your content can leave with you.

Ready for regulated settings

Supports GDPR, Swiss FADP, AI Act and FINMA-regulated contexts · audit trail for an ISO/IEC 42001 programme.

why it matters

Annex III high-risk rules apply from 2 December 2027. Prepare on your timetable, not in a rush.

Neural ICE · included with every AI device

Understands, analyses and delivers.

Neural ICE wraps your chosen model with retrieval, structured understanding, deterministic tools and bounded agents. The model drafts; tools retrieve, calculate and produce files on the device.

Multimodal RAG

Meaning from text and images

Indexes text and page visuals from PDFs and images; returns the passage or page with context.

Grounded extraction

Facts with provenance

Schema-guided facts with provenance and reusable datasets.

Data and KPI production

Computes instead of inventing

Tools query and aggregate tables. KPIs stay reproducible.

Professional deliverables

Editable files, not just chat

PDF, Word, Excel, charts and presentations on the device.

Mail and calendar

Google Workspace and Microsoft 365

Optional mailbox search, reviewed drafts and calendar events. Off by default.

Swiss case law · optional

More than one million decisions, local

More than one million Swiss decisions, searched locally in FR/DE/IT. Packages OpenCaseLaw CC0 + MIT ↗.

On your premises

Share the appliance. Not your knowledge.

Several professionals work on one device; the format sets the ceiling. Desk holds up to 2 users; Workstation up to 8 with one GPU or 16 with two; Station, coming soon, up to 32. Your context and sources stay yours, on site, not a corporate chatbot in the cloud.

The team

Several people, one appliance

One appliance serves several professionals. Each keeps their own work context.

The boundary

Files stay in your organisation

Knowledge and files stay inside your organisation. Premises hardware, not a multi-tenant cloud pool.

On-premises costs more than a shared cloud seat. That premium buys dedicated hardware and clearer organisational boundaries.