Platform / Architecture
Target architecture · in developmentOne operating story. From hardware to workflow.
Use local models first, and only Frontier models when necessary, and approved
Clear boundaries.
Shared operational context.
An infrastructure orchestrator manages clusters, compute, networking, and platform services. An AI orchestrator coordinates models, tools, data, and approvals. Ataraxia’s design connects these layers without treating them as the same thing.
Private execution is the default product direction. Any external model endpoint needs an explicit policy and a clear account of what data can leave the environment.
Connected or fully air-gapped deployment, disconnected lifecycle management and multi-site administration give you the flexibility and security you need in restricted markets.
Responsibilities
Know what each layer owns.
| Layer | Responsibility | Product |
|---|---|---|
| Infrastructure | Cluster provisioning, configuration, health, updates, and recovery | Anvil |
| Workflow development | Guided design, tools, retrieval, testing, and versioning | Atelier |
| Execution | Deploy and run approved workflows with defined resource needs | Atelier runtime on Anvil |
| Model routing | Select an approved endpoint under data, latency, and cost policies | Atelier Gateway |
| Governance | Permissions, approval checkpoints, and execution records | Across the workflow lifecycle |
| Observation | Link execution evidence to evaluations and reviewed changes | Atelier Observe |
The improvement loop
Evidence should lead to action.
The lifecycle pipeline drives continual improvement. The Observation agent is always looking for ways to improve.
Observe execution
Collect sources, tool calls, failures, latency, and resource use.
Evaluate a concern
Compare results against defined graph policies, expected behavior, and human feedback.
Prepare a change
Adjust the workflow graph, data retrieval, model choice, model training, or configuration in a new version.
Review and redeploy
Validate the change, apply policy checks, approve it, and follow its behavior in operation.
Build with us
Start with one useful workflow.
Bring your infrastructure, your operational constraints, and a problem you need solved. Together, we can define a focused pilot and what success should look like.