A hybrid AI platform you can run anywhere

Five modules that work together. Start with one service, keep your data where it belongs, and move workloads between local hardware and the cloud without rewriting your applications.

On-premise or edge Cloud, EU regions Your applications API e Policy engine Local inference a Your data sources c Cloud GPUs b Console d
Fig. 2. Request flow through the platform. Letters refer to the modules below.

Hybrid architecture

Every request enters through the API. The policy engine checks what data it carries and what it may cost, then sends it to local inference or to the cloud. Connectors supply context; the console records the outcome.

  1. Local inference

    Open source: Apple silicon today

    Runs open-weight models on hardware you control. Today that means Apple silicon workstations, through PureDrive, our open-source engine with exact speculative decoding and custom GPU kernels. Servers and edge devices are next on the roadmap.

    PureDrive

    • OpenAI-compatible local server
    • Exact: greedy output is identical with and without speed-ups
    • Long context, up to 128K tokens and beyond

  2. Cloud orchestration

    In development

    Sends a request to the cloud only when a policy allows it: demand peaks, very large models, or tasks with no sensitive data. Designed for NVIDIA GPUs, with the provider and region you choose.

    • Policy-based routing by data class, cost and latency
    • EU regions by default
    • Automatic fallback between local and cloud

  3. Data connectors

    In development

    Reads documents, mailboxes, ticketing systems, CRMs and databases so answers are grounded in your own records, and respects the permissions each system already enforces.

    • Retrieval over files, email and records
    • Per-user access control carried through
    • Data stays in its source system

  4. Console

    In development

    One place to see every request: where it ran, which model answered, what it cost, and how it scored on the quality checks you define. Reviewers can approve or correct answers.

    • Cost and usage per service
    • Quality checks and human review
    • Exportable audit log

  5. API

    In development

    One OpenAI-compatible endpoint for chat, embeddings and tool calling. Existing SDKs and agent frameworks work unchanged, whether the model runs in your rack or in the cloud.

    • Drop-in for OpenAI-style clients
    • Streaming and tool calling
    • Keys and quotas per application

Security and privacy by design

Privacy is the reason the platform runs locally in the first place. Every module is built to these requirements.

  • GDPR first

    Personal data can be pinned to local inference by policy, so it never reaches a third-party model provider.

  • Data in the EU

    Cloud workloads are designed to default to EU regions, including European providers such as Scaleway and OVHcloud.

  • Least privilege

    Connectors inherit source-system permissions; API keys are scoped per application.

  • Full audit trail

    Every request is logged with model, location and outcome, ready for internal review.

  • No training on your data

    Your prompts, documents and outputs are never used to train models.

  • Encryption

    TLS in transit; on-premise deployments use your own storage and keys.

See it on your own use case

Request early access to try the platform on one of your services.

Request early access