Zyrabit Cortex v2.0.0

Sovereign AI Infrastructure for Regulated Environments

Zyrabit Cortex is the local AI operating system for banks, insurance providers, and law firms. Llama-3 inference in air-gapped systems, with zero data leaks and zero cloud dependencies.

Llama-3 inference in under 40 ms on standard corporate hardware.
Zero-trust architecture isolating model network execution (internal: true).
PII scrubbing and audit trails within api-rag before any sensitive data processing.
Docs GitHub
zyrabit@cortex-host ~ zsh

$ _

Sovereign Solutions Catalog

Zyrabit SLM (Free) & Zyrabit Cortex (Enterprise)

Zyrabit offers two complementary tiers. Zyrabit SLM is the free, open-core local inference engine for technical teams. Zyrabit Cortex is the Enterprise layer with operational support, extended security, and certified deployment blueprints for highly regulated environments.

Free · Open-CoreZyrabit SLM

Sovereign Local Inference Engine

Zyrabit SLM is the local inference architecture designed for air-gapped environments. It executes small and specialized models (1B–8B) on standard corporate hardware, keeping the entire data lifecycle within your own infrastructure.

It provides native support for Ollama and local hardware acceleration (NVIDIA CUDA, Apple Metal API, Tenstorrent Grayskull), along with a private RAG pipeline powered by Chroma DB. Zero commercial API dependencies.

  • Lifecycle orchestration via the zyra-up.sh script (install, start, verify, doctor).
  • Internal Docker subnet model-network configured as internal: true, with zero WAN egress.
  • Local active sanitization layer (PII scrubbers) embedded in api-rag before index or query execution.

Start Free

git clone https://github.com/Zyrabit-tech/zyrabit-SLM.git
cd zyrabit-SLM && ./zyra-up.sh install
Enterprise

Enterprise Runtime for Regulated Environments

Zyrabit Cortex is the Enterprise version built upon Zyrabit SLM. Engineered for banks, insurance providers, public sector, and critical industries requiring additional safety guarantees, compliance audits, and production-grade support.

Adds hardware-verified secure enclaves, EFREG optimizations to compress active RAM consumption, and the Gatekeeper FSM state machine to inspect and redact prompt streams in real-time before inference.

  • Native integrations with enterprise Identity Providers (SSO, Active Directory and equivalents).
  • Advanced data retention policies, immutable audit logs, and grounding traceability.
  • Deployment playbooks for air-gapped systems and private VPCs under strict regulatory frameworks.

Enterprise Collaboration Models

  • Guided Enterprise PoC (from £15k) on your own local compute infrastructure.
  • Operational 24/7 SLA-backed support.
  • Private network topology engineering and architectural safety reviews.

Operational Workflow

From Zero to Sovereign Inference in 3 Steps

Zyrabit compresses the complexity of local-first AI infrastructure into three transparent and auditable phases. At the end, you have a corporate inference node completely isolated from the public cloud, running on your own datacenter or dedicated server.

Step 1

Spin Up the Local Stack (Zyrabit SLM)

Installation on standard corporate hardware

Clone the official repository, validate that your server meets GPU/CPU/RAM requirements using our doctor utility, and bring up the complete air-gapped network topology with a single command.

  • Clone from GitHub the official repository.
  • Validate hardware and environment with ./zyra-up.sh doctor.
  • Bring up the complete stack with ./zyra-up.sh install.

CLI (Commands)

git clone https://github.com/Zyrabit-tech/zyrabit-SLM.git
cd zyrabit-SLM
./zyra-up.sh doctor
./zyra-up.sh install
$ git clone https://github.com/Zyrabit-tech/zyrabit-SLM.git
Cloning into 'zyrabit-SLM'...
remote: Enumerating objects: 420, done.

$ cd zyrabit-SLM
$ ./zyra-up.sh doctor
[DR] Checking RAM, CPU and Docker daemon...
[OK] Environment ready for local inference.

$ ./zyra-up.sh install
[UP] Building images and starting stack...
[OK] Zyrabit SLM stack is up and running.
Step 2

Connect Your Knowledge Sources (RAG + Data)

Local indexation, without data ever leaving your server

Mount local volumes containing your organization's internal documents. Configure Chroma DB collections by domain (credit, insurance, legal) and enable PII scrubbing before indexing. Your content never leaves your internal network.

  • Mount secure local volumes with your internal documents (/var/data).
  • Segment collections in Chroma DB by business domain.
  • The api-rag pipeline applies PII scrubbing before indexing.
  • Total control: you define what the model can query and what is restricted.

DOCKER (Volume mount)

docker run -d \
  -v /var/data:/data \
  zyrabitcore/zyrabit-slm:2.0.0
Step 3

Connect Your Interface or Use Ours

Open REST APIs, automation, and production-ready UI

Zyrabit exposes REST APIs and standard Socket.IO channels to integrate your own dashboard, enterprise platform, or automation flows. If you prefer to start immediately, our pre-configured web UI is ready in seconds.

  • REST APIs (/v1/chat, /v1/ingest) and real-time Socket.IO channels.
  • Native automation adapters for n8n, Make and messaging bots.
  • Pre-configured reactive web UI, ready to use with zero extra configuration.
  • Total flexibility: headless, embedded or standalone depending on your architecture.

API QUERY (CURL)

curl -k -X POST https://localhost/v1/chat \
  -H "Content-Type: application/json" \
  -d '{"text":"¿Qué es Zyrabit y cuál es su arquitectura?"}'

Physical Isolation Topology

Zero-Trust Air-Gapped Architecture

Zyrabit Cortex implements network isolation across three logical layers. Knowledge ingestion, contextual grounding, and model inference take place in hermetic containers with zero outbound internet sockets. The model never directly interfaces with the public WAN.

01 · frontend-network (bridge)

The single network border exposed to the host system. Traefik coordinates TLS, rate-limiting, and client authentication. No model is directly accessible; all HTTP interactions traverse this controlled gateway.

  • Local TLS termination and cryptographic negotiation.
  • Rate-limiting rules to prevent abuse or bulk exfiltration.
  • Internal routing directed exclusively toward backend APIs.

02 · backend-network (bridge)

Orchestrates the context grounding and observability pipeline. Hosts api-rag, chroma-db, prometheus, and the gatekeeper-fsm agent, auditing all requests before they reach the execution engine.

  • api-rag: entrypoint for chat/RAG on private corpus.
  • chroma-db: local vector space for ingested documents.
  • prometheus: internal latency metrics and system health.
  • gatekeeper-fsm: active PII and pattern redaction.

03 · model-network (internal: true)

Physical model inference tier (Ollama / Tenstorrent). Provisioned in Docker with internal: true and zero WAN routing. The only ingress allowed is sanitized traffic routed from the backend-network.

  • Model containers do not expose public ports to the host.
  • Zero outbound routing to internet or cloud endpoints.
  • Engineered for air-gapped server grids and datacenters.

Load and Efficiency Metrics

Latency Benchmarks Under Real Scenarios

We publish reproducible benchmarks measured under typical production scenarios. The coordinates displayed here represent real, verifiable performance metrics using our local inference engine and testing scripts.

Local Coding & LLM Benchmarks

ModelHardwareContext / Specsp95 LatencySpeedModeStatus
Qwen-2.5-Coder-7BMac M1 Pro (16 GB)BS=1 • Q4_K_M (4-bit)≈55 ms≈32 tok/s100% localVerified
Llama-3-8B (Cortex)Mac M1 Pro (16 GB)BS=1 • Q4_K_M (4-bit)≈65 ms≈28 tok/s100% localVerified
Qwen-3.6-Coder-7BMac M1 Pro (16 GB)BS=1 • Q4_K_M (Future)≈58 ms≈30 tok/s100% localTesting
Gemma-4-9BMac M1 Pro (16 GB)BS=1 • Q4_K_M (Future)≈85 ms≈18 tok/s100% localTesting
Hardware Baseline: Tests conducted on MacBook Pro M1 Pro (8-Core CPU, 14-Core GPU, 16 GB Unified Memory, 200 GB/s bandwidth).

PII Sanitization & Telemetry

PII Sanitization Pipeline

The active security layer identifies and anonymizes sensitive entities before they hit the model:

secure_agent.py — Sandbox

$ python secure_agent.py "My email is admin@company.co.uk"

# Output: "The email <USER_EMAIL_1> has been processed..."

📊 Telemetry & Metrics

Access real-time system telemetry to audit sovereign resource consumption via Prometheus/Grafana:

zyrabit_token_usage_total

1,420,530

Token Usage

zyrabit_token_latency_ms_per_token

31.4 ms

p95 Latency / Tok

zyrabit_security_hits_total

34 hits

Security Hits
Audit Log Verified

Hermetic Execution Enclaves

100% sovereign
Enclave 01 — Compute Layer

Direct Hardware Accelerator Binding

Direct interface optimization via Apple Metal API and NVIDIA CUDA, enabling sub-40ms execution on standard corporate silicon without API virtualization latency.

Enclave 02 — Data Perimeter

Gatekeeper FSM Stream Interceptor

Strict volatile memory sandbox processing. Prompts are automatically scrubbed and audited before they hit model weights, preventing any PII or customer data leaks.

Enclave 03 — Grounding Scope

Private Isolated Chroma Vector Space

Private corpus context matching executed 100% offline. High-performance semantic indexes run alongside the model engine under the same air-gapped security container.

Zero WAN Network Dependency

High-Security Corporate Case Studies

Where Zyrabit Delivers the Highest Impact

Zyrabit is deployed in environments where the commercial cloud is unfeasible due to data sovereignty mandates, physical security protocols, or critical system dependencies.

Banking

On-premise credit assessment and regulatory compliance

Current State:

Manual ingestion of financial statements and credit bureau histories into local spreadsheets, analyzed manually by senior team members.

Risk:

Exposure of sensitive banking PII to public AI endpoints, and persistent delays when processing flows manually.

Solution:

Zyrabit Cortex is deployed directly onto the bank's local servers. Files are indexed in a private RAG engine, generating risk scores in milliseconds with auditable PII scrubbers compliant with regulators (e.g. CNBV).

Insurance Providers

Policy extraction and contract analysis within secure air-gapped enclaves

Current State:

Decentralized processing of thousands of health and life insurance contracts containing private medical history across isolated shares.

Risk:

Severe regulatory compliance penalties under GDPR or UK-DPA if patient metadata is processed by multi-tenant cloud APIs.

Solution:

Zyrabit SLM and Cortex run natively on local Apple Silicon or x86 servers. Policies are ingested into local Chroma DB systems, enabling private contextual Q&A without any medical bytes traversing the WAN.

Law Firms & Arbitration

RAG search across highly confidential case files and litigation documents

Current State:

Manual inspection of legal precedents, liability clauses, and case timelines across folders and shared network shares.

Risk:

Unacceptable exposure of active litigation strategies to third-party models trained on bulk consumer corpora.

Solution:

Zyrabit SLM compiles semantic indices of historical contracts. Attorneys query precedents privately via local RAG under isolated hardware channels, omitting external WAN interfaces entirely.

Critical Infrastructure & Gov

Offline operational assistance for isolated energy grids and public health

Current State:

Operators at isolated utility grids, military vessels, or clinics consulting thousands of PDF pages, raising downtime risks.

Risk:

Critical operational paralysis during satellite or network outages, or high latency when reliant on remote endpoints.

Solution:

Zyrabit Cortex is orchestrated via Docker on local off-grid servers. Technical manuals and operational protocols are indexed locally, enabling instant cognitive answers without internet.

London Office for Enterprise UK

Local Support & European Regulatory Compliance

Zyrabit Cortex operates through our registered corporate office in London, serving Enterprise systems across the United Kingdom and Europe. We support banks, insurance firms, and public entities in executing local sovereign AI deployments compliant with GDPR and UK-DPA frameworks.

  • Operational engineering and architecture support in European timezones.
  • Regulatory compliance audits and private network topology review.
  • Technical guidance for native on-premise or European private VPC deployments.

Legal Entity

Zyrabit LTD (England & Wales)

Region · UK & Europe

Engineered for highly regulated institutions demanding that patient/banking data and neural networks remain within European jurisdictions, maintaining total infrastructure sovereignty.

We coordinate proofs-of-concept, deployments, and compliance audits in close cooperation with your local engineering and audit leads, aligned to European regulatory standards.

Dev Portal & Assets

Enterprise Sovereign Resources

Access the source files, read structural guides, and pull the latest verified Docker tags. Deploy Zyrabit SLM in isolated enclaves with zero external runtime requirements.

Documentation

Technical Documentation

Step-by-step guides breaking down local installation procedures, multi-tier isolation architectures, Gatekeeper FSM security enclaves, and technical performance validation specs. Built for MLOps, CISOs, and compliance teams.

Source Code

Zyrabit SLM Repository

Fully open-source modular inference engine framework. Houses lifecycle scripts, Docker network files, api-rag configuration, and contribution blueprints for local-first systems.

Deployment

Official Docker Images

Pre-compiled, ready-to-run containers structured for instant edge execution. Contains the core Zyrabit SLM engine and all operational RAG layers needed to boot offline.

CLI (Latest Stable Tag)

docker pull zyrabitcore/zyrabit-slm:2.0.0

Zyrabit Community & Sovereign Ecosystem

Open Collaboration for Local-First AI

Zyrabit is built in close partnership with infrastructure engineers, corporate security units, and local-first AI advocates. Our public channels enable engineers to discuss issues, propose upgrades, and coordinate integrations across the Zyrabit SLM and Cortex stack.

Technical Coordination Channels

GitHub as the Collaboration Hub

We actively collaborate on resolving local orchestration issues, deploying private RAG security patches, and testing specialized hardware drivers. GitHub serves as our primary engineering coordinate for changes and contributions.

  • GitHub Discussions for technical debates and architectural RFCs.
  • GitHub Issues for structured bug reports and capability requests.
  • Pull Requests for contributing directly to the core engine and orchestrators.
Roadmap, Security & Integrations

2026 Roadmap & Responsible Disclosure

We maintain a public roadmap and strict vulnerability disclosure guidelines. We evaluate custom edge connectors and enterprise adapters under clear rules of privacy and mandatory code reviews.

  • 2026 Roadmap focused on offline RAG connectors and specialized hardware drivers.
  • Vulnerability disclosure programs with encrypted channels directly to our security leads.
  • Integration programs for offline vector stores and corporate data sources.

Vulnerability Disclosure

contact@zyrabit.com(PGP optional)

Integrations & Partnerships

contact@zyrabit.com

Technical Specifications · FAQ

Frequently asked technical questions regarding hardware specifications, network isolation, supported model scopes, and Enterprise PoC logistics.