drift-lens-monitor

Your metrics lie. Your embeddings don't.

Quick Start • Why drift-lens? • Methods • CLI • API • Dashboard • Contributing

---

The Problem

Your model's accuracy looks fine — until it doesn't. By the time precision drops, your embeddings have been silently drifting for days. Traditional monitoring watches the wrong signals: accuracy, loss, and latency are lagging indicators. The embedding space is the leading indicator, and nobody's watching it.

drift-lens-monitor detects embedding space drift days before your accuracy drops — with zero infra changes.

Install

pip install drift-lens-monitor

Quick Start

from drift_lens import EmbeddingLogger, DriftDetector

logger = EmbeddingLogger(path="./embeddings", window="1d")
logger.log(embeddings)  # numpy, torch, or list — we handle it

detector = DriftDetector(method="frechet")  # or "mmd" or "topology"
result = detector.compare(baseline_embeddings, current_embeddings)
print(f"Drift score: {result.drift_score:.3f} | Alert: {result.is_drift}")

That's it. Five lines. No API keys, no cloud, no YAML files.

---

Why drift-lens?

	drift-lens	Evidently	Arize
Topology-aware detection	✅ Persistent homology catches cluster merges/splits	❌ Statistical tests only	❌ Statistical tests only
Zero infrastructure	✅ pip install and go — flat files, no DB	⚠️ Requires dashboard server	❌ SaaS platform, needs API keys
Embedding-native	✅ Built specifically for embedding spaces (FED, MMD, Wasserstein)	⚠️ Generic feature drift, bolt-on embedding support	⚠️ General-purpose observability

---

Detection Methods

1. Fréchet Embedding Distance (FED)

The FID of embeddings. Fast, interpretable, great first pass.

• Models embeddings as multivariate Gaussians
• Measures: FED = ‖μ₁ - μ₂‖² + Tr(Σ₁ + Σ₂ - 2·(Σ₁·Σ₂)^½)
• Numerically stabilized with eigenvalue clipping and covariance regularization
• Best for: Quick daily checks, CI/CD gates, dashboards

2. Maximum Mean Discrepancy (MMD)

No Gaussian assumption. Kernel-based. Comes with a p-value.

• RBF kernel with adaptive median-heuristic bandwidth
• Permutation test (200 permutations) for statistical significance
• Detects non-linear distribution shifts FED would miss
• Best for: Rigorous statistical testing, regulatory compliance

3. Topological Drift Detection 🧬

Our moat. No other production drift tool does this.

• Computes persistent homology via Vietoris-Rips filtration (Ripser)
• H₀ features = connected components (cluster count changes)
• H₁ features = loops and holes (structural topology shifts)
• Compares persistence diagrams via Wasserstein distance
• Best for: Detecting cluster merges, splits, and structural collapse

---

CLI Reference

drift-lens ships as a CLI tool. Every command works on parquet snapshot files.

drift-lens compare

Compare two embedding snapshots:

drift-lens compare \
  --baseline ./snapshots/day1.parquet \
  --current  ./snapshots/day14.parquet \
  --method   frechet \
  --threshold 0.3

Flag	Default	Description
--baseline	(required)	Path to baseline snapshot (file or directory)
--current	(required)	Path to current snapshot (file or directory)
--method	frechet	Detection method: frechet, mmd, topology
--threshold	0.3	Alert threshold (0-1)

drift-lens watch

Continuously monitor a snapshot directory:

drift-lens watch \
  --snapshots ./snapshots \
  --method mmd \
  --threshold 0.4 \
  --interval 60

Flag	Default	Description
--snapshots	(required)	Directory to watch for new parquet files
--method	frechet	Detection method
--threshold	0.3	Alert threshold
--interval	60	Polling interval in seconds

drift-lens report

Generate a standalone HTML drift report:

drift-lens report \
  --snapshots ./snapshots \
  --output drift_report.html \
  --method frechet

Flag	Default	Description
--snapshots	(required)	Directory containing snapshot parquet files
--output	drift_report.html	Output HTML file path
--method	frechet	Detection method

drift-lens dashboard

Launch the interactive Streamlit dashboard:

drift-lens dashboard                    # uses built-in demo data
drift-lens dashboard --snapshots ./data # uses your data

Flag	Default	Description
--snapshots	None	Path to snapshots (uses demo data if omitted)

---

Python API

EmbeddingLogger

Log embeddings to disk as time-windowed parquet snapshots:

from drift_lens import EmbeddingLogger

logger = EmbeddingLogger(path="./drift_data", window="1d")  # 1h, 6h, 1d, 7d

# Log from any source — numpy, torch tensors, or plain lists
logger.log(embeddings, metadata={"model": "minilm", "source": "prod"})

# List and load snapshots
files = logger.list_snapshots()
baseline = logger.load_snapshot(files[0])

DriftDetector

Unified interface for all three detection methods:

from drift_lens import DriftDetector

detector = DriftDetector(method="frechet", threshold=0.3)
result = detector.compare(baseline, current)

print(result.drift_score)  # 0.0 - 1.0 (normalized)
print(result.is_drift)     # True/False
print(result.p_value)      # float (MMD only) or None
print(result.details)      # method-specific diagnostics

DriftProjector

Fit-on-baseline / transform-current pattern prevents data leakage:

from drift_lens import DriftProjector

projector = DriftProjector(method="umap", n_components=2)
projector.fit(baseline)
proj_base = projector.transform(baseline)
proj_curr = projector.transform(current)

projector.save("projection.pkl")
projector = DriftProjector.load("projection.pkl")

AlertEngine

Threshold-based alerting with cooldowns and audit trail:

from drift_lens import AlertEngine

engine = AlertEngine(threshold=0.3, cooldown_hours=6, alert_dir="./alerts")
alert = engine.check(result)

print(alert.fired)               # True/False
print(alert.severity)            # low, medium, high, critical
print(alert.recommended_action)  # human-readable next steps

---

Dashboard

The Streamlit dashboard is the demo that sells itself:

# Quick demo with synthetic data
drift-lens dashboard

# or
streamlit run drift_lens/dashboard.py

Features:

• 🔬 Side-by-side UMAP projections (baseline vs. current)
• 📈 Drift score timeline with accuracy overlay
• 🚨 Early warning banner ("DETECTED 4 DAYS BEFORE ACCURACY DROP")
• 📊 All three methods compared in tabs (FED, MMD, Topology)
• 🎯 Severity-based recommended actions
• 📋 Full alert history table

---

How It Works

Your Model                    drift-lens
─────────                    ───────────
embeddings ──┐
             │   EmbeddingLogger     DriftDetector
             ├──→ .log()  ──→ .parquet ──→ .compare()
             │                              │
             │                        ┌─────┴─────┐
             │                        │  drift_score │
             │                        │  is_drift    │
             │                        │  p_value     │
             │                        └─────┬─────┘
             │                              │
             │                        AlertEngine
             │                         .check()
             │                              │
             │                    ┌─────────┴────────┐
             │                    │ 🟢 low            │
             │                    │ 🟡 medium         │
             │                    │ 🟠 high           │
             │                    │ 🔴 critical       │
             │                    └──────────────────┘

---

Contributing

We welcome contributions! See CONTRIBUTING.md for the full guide.

Quick version:

# Fork & clone
git clone https://github.com/YOUR_USERNAME/Drift_lens.git
cd Drift_lens

# Install in dev mode
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Make changes, then PR

---

License

MIT — see LICENSE.

---

Built by Praful Reddy
_{If drift-lens saved you from a production incident, consider giving it a ⭐}