Resources

Open source stack for private AI

Curated projects we track for self-hosted inference, knowledge systems, agent learning, safety, and edge deployment. External links; verify licenses and fit for your environment.

Self-hosted & embedded AI

Run models on your hardware without cloud dependency.

  • Odysseus

    Self-hosted AI stack for local inference and private assistant workflows.

    Why we track it: closest open analog to Sovra's own sovereign-core runtime — useful reference point for local inference architecture.

OCR

Document and image text extraction for on-prem pipelines.

  • Unlimited OCR (Baidu)

    Open OCR tooling for high-volume text recognition from images and scans.

    Why we track it: feeds the document-ingestion side of an on-device RAG pipeline — scanned manuals and photographed documents, processed without leaving the device.

Agent RL & live learning

Reinforcement and online learning bridges for LLM agents.

  • AReaL

    RL bridge for LLM agents — connect training loops to real agent runtimes.

    Why we track it: relevant to command-execution agents that should improve from real on-device usage rather than only static fine-tuning.

  • ODS (Osmantic)

    Osmantic data system for agent-oriented storage and operational learning flows.

    Why we track it: pairs with agent RL to store the operational/episodic data an on-device agent accumulates over time.

Knowledge base & data

Metadata, catalogs, and retrieval beyond vanilla RAG.

  • DataHub

    Metadata platform for data discovery, lineage, and governed knowledge bases.

    Why we track it: the governance layer teams need once they're running Sovra across a fleet of devices and want to track what data feeds which deployment.

  • PixelRAG

    Pixel-oriented retrieval approach aimed at richer grounding than classic chunk RAG.

    Why we track it: a benchmark candidate against Sovra's default chunk-based RAG for image-heavy technical documents.

Security & LLM safety

Red teaming, guardrails, and structured policy via MCP.

  • DeepTeam (Confident AI)

    LLM red-teaming framework for adversarial testing and vulnerability discovery.

    Why we track it: the kind of adversarial testing a command-execution layer needs before it's trusted to trigger real actions.

  • Ontology MCP Server

    MCP server exposing ontologies so agents can reason over typed domain concepts.

    Why we track it: a pattern for exposing structured domain knowledge (not just plain-text chunks) to a validated command layer via MCP.

Healthcare AI

Open research and tooling from Google Health.

  • Google Health (open source)

    Organization hub for health-related open-source models, datasets, and tools.

    Why we track it: a reference point for domain-specific model and dataset work when Sovra hardware is deployed in clinical or health-adjacent settings.

Meeting & local transcription

Privacy-first capture without sending audio to the cloud.

  • Meetily

    Local meeting AI built in Rust with Ollama — transcription and summaries on-device.

    Why we track it: applies the same no-cloud principle Sovra uses for assistant flows to real-time transcription — good template for a fully local meeting-notes add-on.

Knowledge extraction

Structure text into graphs for downstream retrieval and reasoning.

  • Hyper-Extract

    Hypergraph extraction from unstructured text for richer knowledge representation.

    Why we track it: turns ingested documents into a graph instead of flat chunks — potential fidelity upgrade for the knowledge-base tier's RAG.

Local dev & infrastructure

Emulate cloud services on your laptop for private AI CI and testing.

  • floci

    Lightweight local emulator alternative to AWS-style services for dev and integration tests.

    Why we track it: lets teams building on Sovra emulate cloud services locally during dev/CI instead of needing live cloud accounts for integration tests.

Building on private hardware?

Map these tools to your edge or on-prem stack with our hardware configurator and Sovra tiers.

Configure hardware