Skip to main content

HOW IT WORKS

Architecture and deployment

The runtime, Intelligence Packs, drivers, and the two ways to run a Pack inside a customer environment or on the hosted marketplace.


Huitzo helps regulated industries use AI without losing control of data, workflow, or audit records.

AI applications, called Intelligence Packs, run on the Huitzo runtime. Huitzo is model-agnostic, with pre-built Packs designed to produce results in weeks, not months. This page explains the mechanism: what executes, where AI fits, and what evidence each run leaves behind.


Section 1: THE MECHANISM

What happens when a Pack runs

Your own artifacts, made legible to an AI layer that lives where they live.

FIG. 01. THE SELF-HOSTED SYSTEM, IN YOUR NOUNS.

YOUR NETWORK BOUNDARY YOUR DOCUMENTS YOUR POLICIES YOUR TEAM INTELLIGENCE PACKS DETERMINISTIC LAYER YOUR INFRASTRUCTURE ONE GATE ANY MODEL YOUR CHOICE OF PROVIDER, HOSTED OR SELF-HOSTED PROMPTS LEAVE YOUR DATA STAYS YOUR NETWORK BOUNDARY YOUR DOCUMENTS YOUR POLICIES YOUR TEAM INTELLIGENCE PACKS DETERMINISTIC LAYER YOUR INFRASTRUCTURE ONE GATE ANY MODEL YOUR CHOICE OF PROVIDER, HOSTED OR SELF-HOSTED PROMPTS LEAVE YOUR DATA STAYS
Self-hosted: your documents, your policies, and your team feed the deterministic layer on your infrastructure. Any model, through one gate.

Deterministic Python executes

An Intelligence Pack is business logic written as code your team can read, test, and audit. The code owns the decisions and runs the same way every time.

AI is invoked only where it adds value, bounded

Where a step genuinely needs a model (classify, extract, summarize) the Pack calls one through a declared driver, either a self-hosted model inside your boundary or a hosted provider, and the record states which. AI augments the process; it does not own it.

Every run is recorded

Every run is recorded by the runtime itself: the command, its status and timing, and the Pack's governance level from its Policy Card. The v0.1 draft record format adds what the AI step saw, what was approved and what stayed inside the boundary, signed so a third party can verify it.

Deterministic means the control flow is fixed and testable: the same input runs the same coded path every time, each AI step is bounded, versioned and measured, and the runtime records what ran. Not certainty of output; reproducibility of process.

Every Intelligence Pack ships with a machine-readable Policy Card declaring allowed actions, data-handling scope, escalation requirements, and logging obligations, and the platform logs every action for audit. In self-hosted mode, Packs never phone home; that is enforced by architecture, not policy.


Section 2: THE DIFFERENTIATOR

Observers prove what crossed the wire. Huitzo proves what your application did.

Most AI assurance tools sit beside the application and attest to the calls it made. Huitzo is the runtime the application actually runs on, so the evidence of what happened is produced by the execution itself.


Section 3: THE RUNTIME

The runtime

Like the Linux kernel hosts processes, the Huitzo runtime, Huitzo Hub, hosts Intelligence Packs. It runs on Linux: inside the customer environment when self-hosted, or operated by Huitzo on the hosted marketplace.

YOUR NETWORK BOUNDARY YOUR DATA POSTGRES FILES YOUR INFRASTRUCTURE HUITZO RUNTIME ONE INTERFACE INTELLIGENCE PACKS INTAKE REVIEW AUDIT DRIVER MODEL PROVIDERS OPENAI ANTHROPIC PROMPTS LEAVE YOUR DATA STAYS YOUR NETWORK BOUNDARY YOUR DATA POSTGRES FILES YOUR INFRASTRUCTURE HUITZO RUNTIME ONE INTERFACE INTELLIGENCE PACKS INTAKE REVIEW AUDIT DRIVER MODEL PROVIDERS OPENAI ANTHROPIC PROMPTS LEAVE YOUR DATA STAYS
Self-hosted view: the runtime on your infrastructure, with every Pack running inside it.

Section 4: INTELLIGENCE PACKS

Intelligence Packs

An Intelligence Pack is an AI application: the unit that gets deployed, versioned, and audited. Like a containerized application on a Linux host, but with one shared interface to models, storage, and drivers. A deployment starts from an existing Pack where one fits rather than a blank repository. Where nothing in the library fits, a team builds a custom Pack against the same interface, and it runs on the same runtime inside the same boundary.

YOUR NETWORK BOUNDARY YOUR DATA POSTGRES FILES YOUR INFRASTRUCTURE HUITZO RUNTIME ONE INTERFACE INTELLIGENCE PACKS INTAKE REVIEW AUDIT DRIVER MODEL PROVIDERS OPENAI ANTHROPIC PROMPTS LEAVE YOUR DATA STAYS YOUR NETWORK BOUNDARY YOUR DATA POSTGRES FILES YOUR INFRASTRUCTURE HUITZO RUNTIME ONE INTERFACE INTELLIGENCE PACKS INTAKE REVIEW AUDIT DRIVER MODEL PROVIDERS OPENAI ANTHROPIC PROMPTS LEAVE YOUR DATA STAYS
Self-hosted view: Packs slot into the runtime the way applications install on an operating system.

Section 5: DRIVERS

Drivers

Drivers extend Huitzo to external and internal systems, the way Linux device drivers extend the kernel. Standard drivers ship for OpenAI, Anthropic, and S3-compatible storage. Custom drivers connect the runtime to databases and proprietary systems.

YOUR NETWORK BOUNDARY YOUR DATA POSTGRES FILES YOUR INFRASTRUCTURE HUITZO RUNTIME ONE INTERFACE INTELLIGENCE PACKS INTAKE REVIEW AUDIT DRIVER MODEL PROVIDERS OPENAI ANTHROPIC PROMPTS LEAVE YOUR DATA STAYS YOUR NETWORK BOUNDARY YOUR DATA POSTGRES FILES YOUR INFRASTRUCTURE HUITZO RUNTIME ONE INTERFACE INTELLIGENCE PACKS INTAKE REVIEW AUDIT DRIVER MODEL PROVIDERS OPENAI ANTHROPIC PROMPTS LEAVE YOUR DATA STAYS
Self-hosted view: drivers connect the runtime to models and external systems, and the record states what crossed the boundary.

Section 6: DEPLOYMENT

Two ways to run a Pack

Security

You choose where it runs. Self-hosted, your data never leaves your environment: not in transit, not in processing, not in logs.

On the hosted marketplace, your Pack, its data and its evidence live in your isolated tenant; the only thing that leaves it is the bounded model call, to the provider you chose, and we never train on your data. We tell every buyer which mode they are getting, where the model calls go, and what leaves the tenant, before they sign.

Self-hosted (inside your environment)

The runtime, the Packs, the data and the audit records stay inside infrastructure you control. Nothing is sent to Huitzo, and that includes logs. With a self-hosted model, no prompt leaves your network; with a hosted model provider, only the declared prompts cross, and the record states what crossed.

Hosted marketplace (huitzo.ai)

On the hosted marketplace, your Pack, its data and its evidence live in your isolated tenant; the only thing that leaves it is the bounded model call, to the provider you chose, and we never train on your data.

SELF-HOSTED PATH. FIVE STEPS.

  1. 06.1

    Provision the runtime host

    Stand up a Linux host inside the customer environment. The runtime needs network access to the chosen model provider, or to a self-hosted model, and to the systems the Intelligence Pack will integrate with.

  2. 06.2

    Install the Huitzo runtime

    Run the Huitzo install script. The runtime and the services it depends on run on hosts you control. No customer data leaves the host during installation.

  3. 06.3

    Configure drivers

    Drivers connect the runtime to model providers, object storage, and internal systems. Standard drivers ship out of the box; custom drivers extend the runtime to anything proprietary.

  4. 06.4

    Deploy an Intelligence Pack

    An Intelligence Pack is the application unit that runs on the runtime. Build locally with the Python SDK, then deploy to the runtime.

  5. 06.5

    Open the dashboard

    The dashboard that ships with the runtime is where your team monitors and manages deployed Intelligence Packs.


HOSTED PATH.

  1. A partner publishes the Intelligence Pack to the huitzo.ai marketplace.
  2. The client runs the Pack inside its own isolated tenant.
  3. The partner stays the commercial and local face of the relationship; Huitzo operates the platform.

Section 7: ENGAGEMENT

From conversation to production

Pre-built Intelligence Packs are designed to produce results in weeks rather than the 12-18 months a fully custom internal build may require, and every Pack runs where the customer decides.

Conversation

A working session with the founder. Your environment, your processes, and whether an Intelligence Pack fits. No deck.

Scoped evaluation

One process becomes an Intelligence Pack, on huitzo.ai or inside your environment, recorded from the first run so you can judge it.

Production

The Pack goes live where your data lives, and every run keeps producing its own evidence.


Section 8: QUESTIONS

How it works: questions

What gets installed on customer infrastructure?

On the self-hosted install path: the Huitzo runtime and the services it depends on, on Linux hosts the customer controls, plus the Intelligence Packs the customer chooses to deploy. There is no Huitzo-managed cloud component on the data path for a self-hosted install. On the hosted marketplace, nothing installs on customer infrastructure; Huitzo operates the platform, and the Pack runs inside the customer's isolated tenant.

Where do AI models run?

Huitzo is model-agnostic in both delivery modes. The customer chooses: a hosted provider (OpenAI, Anthropic) reached over the network, or a self-hosted model on customer infrastructure for fully air-gapped deployments (self-hosted mode only).

How long does deployment typically take?

Pre-built Intelligence Packs are designed to produce results in weeks rather than the 12-18 months a fully custom internal build may require, and every Pack runs where the customer decides. Air-gapped self-hosted environments can take longer because of additional integration scope.

How do Intelligence Packs differ from containers?

An Intelligence Pack is a Huitzo-runtime application, like an ELF binary on Linux. It runs on top of the runtime, which provides the shared interface to models, storage, and drivers. A container is a packaging format; an Intelligence Pack is an application unit.


Section 9: CLOSING

AI operating system for regulated industries

Simple by design. Infinitely scalable. Runs where your data lives.

Product access
huitzo.ai, the hosted marketplace, is open by invitation to teams evaluating the product.

Or email [email protected] directly.