Externalized model inference for the Nvisy runtime.
Hosts each model as a standalone HTTP service so the runtime calls inference over the network instead of loading ML runtimes in-process. Each model category ships as its own independently scalable, independently deployable service behind a stable wire contract.
- Contract-first, swappable engines: services implement a versioned wire contract, not a specific model — the engine behind each contract can be replaced without touching the runtime
- Model categories: detection OCR (text + word-level geometry), vision-language OCR (high-accuracy transcription and layout), and named-entity recognition over a shared entity taxonomy
- Layered OCR: traditional OCR provides precise geometry while an optional GPU vision-language service refines text accuracy; the runtime reconciles the two
- Independent services: each model runs as its own image with independent scaling and failure domains, and any service can be opted out of
- Bring your own inference: any service that reproduces the wire contract is a drop-in replacement, including self-hosted or custom models and weights
- Lockstep versioning: inference releases track the runtime version so the contract stays in sync
The fastest way to get started is with Nvisy Cloud.
For self-hosted deployments, the services are published as container images and
run alongside the Nvisy runtime; see
docs/ for the wire contract and deployment guidance.
See docs/ for architecture, contract, and API documentation.
See CHANGELOG.md for release notes and version history.
Apache 2.0 License, see LICENSE.txt
- Documentation: docs.nvisy.com
- Issues: GitHub Issues
- Email: support@nvisy.com