ml-reliability

Early-warning research benchmark for synthetic-data-induced model degradation using cross-dataset drift metrics, warning lead time, and statistical validation.

ByteStack Labs marketplace for Claude Code. Open reliability skills that audit AI which passes evaluation but fails in production. Every number reproducible.

The missing CI for your ML supply chain: reads DataHub's column-level lineage and ML metadata to catch silent data-to-model failures

TrainKeeper is a minimal-decision, high-signal toolkit for building reproducible, debuggable, and efficient ML training systems. It adds guardrails inside training loops without replacing your existing stack.

This project is a production-grade autonomous control system designed to maintain machine learning model integrity through a closed-loop Detect → Diagnose → Decide → Act → Explain cycle. Unlike traditional monitoring that requires slow human intervention, SHMLP autonomously identifies data drift, concept shift, and inference anomalies to execute

PyPLTool : Runtime trust layer for machine learning systems. Detects drift, uncertainty, and reliability risks in production ML models.

AI experimental forensics and reliability laboratory (early development): reproducible evaluation, failure analysis, fault injection, provenance, and longitudinal reliability research.

Three-layer RAG system that retrieves, diagnoses, and evaluates engineering post-mortems — with the evaluation framework as the core, not an afterthought.

Leakage-aware, synthetic-only reliability evaluation for rating-proxy review text classification.