Platform
Vareon has no proprietary data. You bring your own. ARDA trains per-run on your data — no massive pre-training datasets required. Deploy cloud-hosted, self-hosted, or air-gapped.
Bring Your Own Data
Unlike large-scale ML systems that require millions of samples, ARDA discovers governing equations and causal structures from small datasets. The preflight system requires as few as 8 total timesteps across all episodes. Resource planning automatically adjusts algorithm parameters for sparse data.
Vareon has no proprietary data and never will. Your data stays yours. ARDA trains per-run on the data you provide — there is no pre-trained foundation model that needs fine-tuning.
Key differentiator
Upload episodes via POST /v1/data/upload with structured JSON payloads. Profile data, list datasets, and manage uploads programmatically.
client.upload_episodes(), client.profile_data(), client.generate_data() — typed methods with full IDE support.
Upload, profile, and generate data through Model Context Protocol tools. Agents can provision data without writing HTTP calls.
POST observations to /v1/streams/{id}/ingest as they arrive. Server-sent events for live status. Automatic trigger for discovery runs.
Data Model
All data enters ARDA as episodes — structured observations over time. Only timestamps and observations are required. Spatial, relational, hierarchical, and causal modalities are automatically detected when present.
Core
Spatial
Relational
Hierarchical
Causal / CDE
Quality
Supported input formats: JSON episodes (native), CSV, Parquet, and HDF5 (via SDK conversion). Molecular dynamics adapter payloads are automatically normalized.
Compute
ARDA estimates compute requirements from your data profile and discovery mode. Resource planning is automatic — it selects hardware, sets runtime budgets, and adjusts algorithm parameters without manual tuning.
Vareon provisions and manages GPU compute. Automatic hardware selection based on data profile and discovery mode. Pay for gpu_seconds consumed. No infrastructure to manage.
Deploy ARDA on your own infrastructure. License fee only — bring your own GPUs. Full control over data residency and network topology.
Fully isolated deployment with offline license activation. No outbound network required. Suitable for classified environments, defense, and regulated industries.
Hardware selection
CPU or GPU, automatic
Runtime budget
Estimated per-run, capped
Concurrency
Multiple runs in parallel
Billing
gpu_seconds consumed
Bring Your Own Agent
ARDA's entire API surface is designed for programmatic invocation. Bring your own AI agent — Claude, GPT, a custom research agent, or any system that can make HTTP calls — and have it orchestrate ARDA as its discovery engine.
A BYO Scientist session gives your external agent operator-level access: it can formulate hypotheses, design experiments, launch runs, and review results through ARDA's structured task system.
Agent capabilities
Integration endpoint
POST /v1/agents/sessions/byo/scientist
{
"project_id": "...",
"objective": "Discover governing equations...",
"dataset_id": "...",
"provider_id": "..."
}
Your agent receives structured tasks, claims its work items, executes discovery runs through the full API, and submits results. ARDA handles governance, provenance, and negative controls automatically.
Bring Your Own LLM
ARDA's agent cognition layer supports any OpenAI-compatible model endpoint. Bring your own LLM provider for agent planning, hypothesis generation, and result interpretation — use your existing enterprise agreements and model preferences.
Configure your provider once with a base URL, default model, and API key. ARDA's agent system routes cognition calls to your endpoint using the standard chat completions protocol.
Configuration
How it works
Contact us to discuss your deployment requirements — cloud-hosted, self-hosted, or air-gapped.