modelpulse-sdk 0.2.0

ML model monitoring + LLM observability — one decorator, live drift detection, AI diagnosis, token cost tracking

modelpulse-sdk

ML model monitoring for developers. One decorator. Live drift detection, AI diagnosis, and instant alerts — in 60 seconds.

---

What is ModelPulse?

ModelPulse monitors your ML models in production. It detects when your model's input data starts looking different from what it was trained on (data drift), generates AI-powered plain-English explanations of what changed, and fires Slack/email alerts — automatically.

No MLOps team required. No schema design. No enterprise sales call.

---

Installation

pip install modelpulse-sdk

---

Quickstart

import modelpulse

# Initialize once at startup
modelpulse.init(
    api_key="mp_live_xxxx",       # from dashboard → model → API Keys
    model_id="your-model-uuid",   # from dashboard
)

# Wrap your predict function with one decorator
@modelpulse.monitor
def predict(features: dict):
    return my_model.predict([features])[0]

# That's it — ModelPulse logs every prediction automatically
result = predict({"age": 34, "income": 72000, "tenure_months": 12})

ModelPulse will:

• Log every prediction (inputs, outputs, confidence, latency)
• Compute drift hourly against your baseline distribution
• Generate an AI diagnosis when drift is detected
• Fire Slack/email alerts with the diagnosis included

---

How it works

1. Batching

Predictions are queued in memory and flushed every 10 seconds or when 50 accumulate — whichever comes first. This means zero blocking on your model's critical path.

2. Non-blocking

The SDK uses a background thread. Your predict() function adds less than 1ms overhead.

3. Resilient

Network failures are silently dropped. The SDK never raises an exception that could crash your application.

---

Advanced usage

Return confidence scores

@modelpulse.monitor
def predict(features: dict):
    probs = model.predict_proba([features])[0]
    # Return (prediction, confidence) tuple
    return probs.argmax(), float(probs.max())

Custom API URL (self-hosted)

modelpulse.init(
    api_key="mp_live_xxxx",
    model_id="your-model-uuid",
    api_url="https://your-own-api.com/api/v1",
)

Works with any framework

# scikit-learn
@modelpulse.monitor
def predict(features):
    return sklearn_model.predict([list(features.values())])[0]

# XGBoost
@modelpulse.monitor
def predict(features):
    dmatrix = xgb.DMatrix([list(features.values())])
    return float(xgb_model.predict(dmatrix)[0])

# PyTorch
@modelpulse.monitor
def predict(features):
    tensor = torch.tensor(list(features.values()))
    return torch_model(tensor).item()

---

Drift Detection

ModelPulse uses two statistical tests under the hood:

Test	What it measures
PSI (Population Stability Index)	Magnitude of distribution shift
KS Test (Kolmogorov-Smirnov)	Whether two distributions are different

PSI Score	Health Status
< 0.10	🟢 Healthy — no significant drift
0.10 – 0.20	🟡 Warning — moderate drift
> 0.20	🔴 Critical — significant drift

You never need to configure these thresholds manually — ModelPulse computes them automatically from your first 100 predictions.

---

Dashboard

After integrating the SDK, view your model's health at:

👉 https://frontend-ruddy-sigma-92.vercel.app

The dashboard shows:

• Feature drift scores with visual progress bars
• AI-generated plain-English diagnosis (powered by Claude)
• Prediction volume and latency trends
• Alert history

---

Configuration options

modelpulse.init(
    api_key="mp_live_xxxx",      # required — your API key
    model_id="uuid",             # required — your model ID
    api_url="https://...",       # optional — default: ModelPulse cloud
    timeout=5,                   # optional — HTTP timeout in seconds
    batch_size=50,               # optional — flush after N predictions
    async_mode=True,             # optional — fire-and-forget mode
)

---

Links

• 🌐 Website: https://landing-kappa-lilac.vercel.app
• 📊 Dashboard: https://frontend-ruddy-sigma-92.vercel.app
• 📁 GitHub: https://github.com/GPREETHAMSAXON/Model_Pulse
• 🐛 Issues: https://github.com/GPREETHAMSAXON/Model_Pulse/issues

---

License

MIT License — see LICENSE for details.

---

Built by Saxon Preetham · Visakhapatnam, India · 2026