Using SKaiNET — Audience and Scope

Audience: people building apps with SKaiNET. These pages document how to call SKaiNET as a library — pulling artifacts from a BOM, constructing tensors, running a forward pass, training a model. If you are modifying SKaiNET itself — adding kernels, running benchmarks, operating CI — see Contributing.

What this section covers

The Using SKaiNET section is for the engineer who:

Adds SKaiNET to a Gradle / Maven build.
Builds tensors, runs forward passes, trains small models.
Loads pre-trained weights (GGUF, SafeTensors, ONNX).
Targets StableHLO / IREE for cross-platform deployment.
Exports supported static MLPs to Minerva secure MCU bundles.
Reads the operator catalog to discover what’s supported on which backend.

Inside, content follows the Diátaxis / Divio framework:

Sub-section	What it answers
Tutorials	Learning-oriented walkthroughs. Start here if you are new to SKaiNET.
How-to guides	Task-oriented recipes — "I need to load a GGUF model", "I need to train from Java".
Reference	Information-oriented lookup: architecture, operator catalog, coverage matrix, Dokka API.
Explanation	Understanding-oriented background: theory, performance notes, design rationale.

Sub-section

What it answers

Tutorials

Learning-oriented walkthroughs. Start here if you are new to SKaiNET.

How-to guides

Task-oriented recipes — "I need to load a GGUF model", "I need to train from Java".

Reference

Information-oriented lookup: architecture, operator catalog, coverage matrix, Dokka API.

Explanation

Understanding-oriented background: theory, performance notes, design rationale.

Using SKaiNET from Java

SKaiNET is a Kotlin Multiplatform library; the JVM artifact is consumable from Java but the idioms differ (builder factories instead of extension functions, no operator overloading). The following pages target Java specifically:

Java getting started — set up a Maven / Gradle project, run a forward pass.
Train a model from Java — assemble a training loop using the Java surface.

The rest of this section is written in Kotlin. The concepts transfer directly; only the syntax differs.

Export and deployment paths

StableHLO getting started — lower graphs to portable MLIR for IREE-compatible compiler flows.
Minerva getting started — export a tiny static MLP to a secure MCU bundle.
Generate C for Arduino — generate standalone C99 for small-device deployment without libminerva.

Working with images and image-shaped tensors

If you are working with preprocessing pipelines or image-shaped tensors, start with Image and data API for the skainet-io-image, skainet-data-transform, and skainet-data-media layers.

LLM-specific Java runtimes (Llama, Gemma, Qwen, BERT) moved to the sibling SKaiNET-transformers repository in 2026.

What this section deliberately does not cover

How SKaiNET is built, benchmarked, or released. See Contributing.
Commit conventions, branch policy, issue reports. Those live in repo-root files (CONTRIBUTING.md, GITFLOW.adoc, CHANGELOG.md, FAQ.md) and intentionally aren’t duplicated here.