kailash-ml

What problem does it solve?

Kailash ML solves the complexity of building reliable end-to-end machine learning lifecycles—from feature ingestion and training to registration, ONNX serving, and drift monitoring—while enforcing safety and consistency constraints that prevent common production failures.

Core Features & Use Cases

• Polars-native ML lifecycle: keep data in polars end-to-end and only convert at framework boundaries, reducing mismatches and leakage risks.
• Engine-first API surface (km.*): provides a canonical async workflow for tracking, training, registering, diagnosing, watching drift, resuming, and serving.
• Production-grade governance: includes model registry lifecycle (staging → shadow → production → archived), SHA256 integrity checks, GDPR erasure controls, and agent guardrails for AutoML.
• Serving + drift integration: serve production models via Nexus with ONNX-first runtime options, plus drift monitoring triggers and retraining hooks.
• ONNX export + verification: export sklearn/lightGBM/PyTorch models to ONNX and verify outputs before deployment for cross-language compatibility.

Quick Start

Train and register a production-ready model by running an async pipeline that tracks a run, trains from a polars DataFrame, then registers and serves the resulting artifact using the engine-first km.* verbs.