Product · Self-serve · Coming September

Helios Factory.
Compile your own world model.

The same factory we use, self-serve. Bring your ontology, data, action space, and governance. Compile, simulate, and embed the result inside your own systems.

Or have us build it

Built with and for high-stakes engineering teams

TIER-1 BANK

SOVEREIGN WEALTH FUND

EUROPEAN INSURER

PUBLIC ASSET STEWARD

NATIONAL UTILITY

MINISTRY OF FINANCE

GLOBAL REINSURER

REAL ESTATE GROUP

TELECOM OPERATOR

HEALTHCARE NETWORK

TIER-1 BANK

SOVEREIGN WEALTH FUND

EUROPEAN INSURER

PUBLIC ASSET STEWARD

NATIONAL UTILITY

MINISTRY OF FINANCE

GLOBAL REINSURER

REAL ESTATE GROUP

TELECOM OPERATOR

HEALTHCARE NETWORK

TIER-1 BANK

SOVEREIGN WEALTH FUND

EUROPEAN INSURER

PUBLIC ASSET STEWARD

NATIONAL UTILITY

MINISTRY OF FINANCE

GLOBAL REINSURER

REAL ESTATE GROUP

TELECOM OPERATOR

HEALTHCARE NETWORK

TIER-1 BANK

SOVEREIGN WEALTH FUND

EUROPEAN INSURER

PUBLIC ASSET STEWARD

NATIONAL UTILITY

MINISTRY OF FINANCE

GLOBAL REINSURER

REAL ESTATE GROUP

TELECOM OPERATOR

HEALTHCARE NETWORK

What it is

The factory, in your hands. Run it where your systems run.

The Factory is a composable runtime for teams who want full control. You compose the same thirteen substrate components we use, ontology compiler, data wiring, eight reasoning layers, governance, provenance, and deploy the resulting executable model inside your own infrastructure.

Capabilities · Composable. Inspectable. Yours.

The same factory we use, in your hands.

Bring your own ontology

Compose your typed graph in YAML, or extend one of our open templates for banking, healthcare, energy.

Bring your own data

Postgres, BigQuery, Snowflake, Kafka, S3. The factory wires provenance automatically.

Compile and simulate

The same eight reasoning layers we use. Rehearse millions of variants before committing to a path.

Embed in your systems

Self-host. VPC-deploy. Air-gapped. The factory respects whatever boundary your governance demands.

How the factory works

Six steps. From spec to runtime.

· 01

Compose your ontology

Author a typed graph of your domain: entities, relations, time, provenance. Versioned. Reviewable. Diffable.

· 02

Wire your data

Bind every variable to its source. Provenance is automatic, every signal carries who, when, and at what fidelity.

· 03

Define the action space

Codify what the model can recommend and how each recommendation is governed before it leaves the runtime.

· 04

Compile & simulate

The factory compiles your spec into an executable world model. Rehearse millions of variants before committing.

· 05

Embed in your systems

Deploy as a service inside your infrastructure. SDKs for Python, TypeScript. APIs for the rest.

· 06

Close the loop

Outcomes feed back into the substrate. The model learns from the world it shaped, auditably, governably.

Hands-on · What you can ship with it

Move from scripts to a reasoning runtime.

Data Platform Team

Compose an executable world model in a Python repo

Author your ontology + action space, wire your warehouse, get a runtime your engineers can review like any other service.

Outcome

From idea to first decision: 3 weeks

Quant Engineering

Replace ad-hoc Python scripts with a governed reasoning runtime

Same code shape, but every decision now carries provenance, refusal, and audit. No more shadow systems.

Outcome

Audit prep: 6 weeks → 4 days

Risk Engineering

Codify your action space as YAML, enforce at runtime

Every recommendation that leaves the runtime carries the policy it satisfied. Refusals are first-class.

Outcome

Policy drift: zero in 6 months

Platform · Insurance

Self-host inside an air-gapped VPC

Deploy a complete world model behind your firewall. No model weights leave. No prompts leave.

Outcome

First production decision: 11 weeks

Public Sector IT

Compose a model from a federated registry

Wire two registries, declare your safety properties, watch SMT certify or counter-example them.

Outcome

Federation cleared first audit

ML Platform

Ship a model that learns from its own outcomes

Outcome ledger + scoring rules feed back into the substrate. The model improves the way the institution measures.

Outcome

Calibration drift: −62% in 4 months

The factory, in numbers

Substrate components

Reasoning layers

Steps in the learning loop

100%

Inspectable, end to end

A glimpse

Compose your model in a few lines.

helios · brain.py

from helios import Factory, Ontology, Data, ActionSpace

# 1. Compose
brain = Factory()
brain.add(Ontology.load("./ontology/banking.helios"))
brain.add(Data.from_postgres(conn, provenance=True))
brain.add(ActionSpace.from_yaml("./actions/credit.yaml"))

# 2. Compile
model = brain.compile(governance="eu-ai-act")

# 3. Rehearse
trace = model.simulate(
    action="tighten_credit_policy",
    horizon="24w",
    n_variants=1924
)

# 4. Embed
model.serve(port=8080, infra="vpc")