Overview

Driftlite detects when ML models start behaving differently in production—before accuracy tanks. Uses simple, interpretable statistics instead of complex monitoring systems.

Accompanies the preprint: "Lightweight Monitoring of AI System Drift Using Interpretable Statistical Signals"

---

Research Questions

1. Detection Sensitivity — How reliably can PSI and KS-Test identify distribution shifts?
2. Performance Correlation — How does detected drift relate to accuracy degradation?

---

Methods

• PSI (Population Stability Index) — Measures feature distribution shift
• KS-Test — Non-parametric comparison of probability distributions

Both methods are zero-dependency (just SciPy/NumPy) and model-agnostic.

---

Features

• Realistic drift simulation (noise, label corruption)
• Automated Markdown reporting with distribution plots
• Works with any Scikit-learn model

---

Tech Stack

Python SciPy NumPy Scikit-learn Research