ai-critic 1.2.0

Fast AI evaluator for scikit-learn models

ai-critic 🧠: The Quality Gate for Machine Learning Models

ai-critic is a specialized decision-making tool designed to audit the reliability and readiness for deployment of scikit-learn, PyTorch, and TensorFlow models.

Instead of merely measuring performance (accuracy, F1 score), ai-critic acts as a Quality Gate, actively probing the model to uncover hidden risks that commonly cause production failures — such as data leakage, structural overfitting, and fragility under noise.

ai-critic does not ask “How good is this model?” It asks “Can this model be trusted?”

---

🚀 Getting Started (The Basics)

This section is ideal for beginners who need a fast and reliable verdict on a trained model.

Installation

Install directly from PyPI:

pip install ai-critic

---

The Quick Verdict

With just a few lines of code, you obtain an executive-level assessment and a deployment recommendation.

from ai_critic import AICritic
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import make_classification

# 1. Prepare data and model
X, y = make_classification(n_samples=1000, n_features=20, random_state=42)
model = RandomForestClassifier(max_depth=5, random_state=42)

# 2. Initialize the Critic
critic = AICritic(model, X, y)

# 3. Run the audit (executive mode)
report = critic.evaluate(view="executive")

print(f"Verdict: {report['verdict']}")
print(f"Risk Level: {report['risk_level']}")
print(f"Main Reason: {report['main_reason']}")

Expected Output (example):

Verdict: ⚠️ Risky
Risk Level: medium
Main Reason: Structural or robustness-related risks detected.

This output is intentionally conservative. If ai-critic recommends deployment, it means meaningful risks were not detected.

---

💡 Understanding the Critique (The Intermediary)

For data scientists who want to understand why the model received a given verdict and how to improve it.

---

The Four Pillars of the Audit

ai-critic evaluates models across four independent risk dimensions:

Pillar	Main Risk Detected	Internal Module
📊 Data Integrity	Target Leakage & Correlation Artifacts	evaluators.data
🧠 Model Structure	Over-complexity & Misconfiguration	evaluators.config
📈 Performance	Suspicious CV or Learning Curves	evaluators.performance
🧪 Robustness	Sensitivity to Noise	evaluators.robustness

Each pillar contributes signals used later in the deployment gate.

---

Full Technical & Visual Analysis

To access all internal diagnostics, including plots and recommendations, use view="all".

full_report = critic.evaluate(view="all", plot=True)

technical_summary = full_report["technical"]

print("\n--- Key Risks Detected ---")
for i, risk in enumerate(technical_summary["key_risks"], start=1):
    print(f"{i}. {risk}")

print("\n--- Recommendations ---")
for rec in technical_summary["recommendations"]:
    print(f"- {rec}")

Generated plots may include:

• Feature correlation heatmaps
• Learning curves
• Robustness degradation charts

---

Robustness Test (Noise Injection)

A model that collapses under small perturbations is not production-safe.

robustness = full_report["details"]["robustness"]

print("\n--- Robustness Analysis ---")
print(f"Original CV Score: {robustness['cv_score_original']:.4f}")
print(f"Noisy CV Score: {robustness['cv_score_noisy']:.4f}")
print(f"Performance Drop: {robustness['performance_drop']:.4f}")
print(f"Verdict: {robustness['verdict']}")

Possible Verdicts:

• stable → acceptable degradation
• fragile → high sensitivity to noise
• misleading → performance likely inflated by leakage

---

⚙️ Integration and Governance (The Advanced)

This section targets MLOps engineers, architects, and teams operating automated pipelines.

---

Multi-Framework Support

ai-critic 1.0+ supports models from multiple frameworks with the same API:

# PyTorch Example
import torch
import torch.nn as nn
from ai_critic import AICritic

X = torch.randn(1000, 20)
y = torch.randint(0, 2, (1000,))

model = nn.Sequential(
    nn.Linear(20, 32),
    nn.ReLU(),
    nn.Linear(32, 2)
)

critic = AICritic(model, X, y, framework="torch", adapter_kwargs={"epochs":5, "batch_size":64})
report = critic.evaluate(view="executive")
print(report)

# TensorFlow Example
import tensorflow as tf

model = tf.keras.Sequential([
    tf.keras.layers.Dense(32, activation="relu", input_shape=(20,)),
    tf.keras.layers.Dense(2)
])
critic = AICritic(model, X.numpy(), y.numpy(), framework="tensorflow", adapter_kwargs={"epochs":5})
report = critic.evaluate(view="executive")
print(report)

No need to rewrite evaluation code — one Critic API works for sklearn, PyTorch, or TensorFlow.

---

The Deployment Gate (deploy_decision)

The deploy_decision() method aggregates all detected risks and produces a final gate decision.

decision = critic.deploy_decision()

if decision["deploy"]:
    print("✅ Deployment Approved")
else:
    print("❌ Deployment Blocked")

print(f"Risk Level: {decision['risk_level']}")
print(f"Confidence Score: {decision['confidence']:.2f}")

print("\nBlocking Issues:")
for issue in decision["blocking_issues"]:
    print(f"- {issue}")

Conceptual model:

• Hard Blockers → deployment denied
• Soft Blockers → deployment discouraged
• Confidence Score (0–1) → heuristic trust level

---

Modes & Views (API Design)

The evaluate() method supports multiple modes via the view parameter:

View	Description
"executive"	High-level verdict (non-technical)
"technical"	Risks & recommendations
"details"	Raw evaluator outputs
"all"	Complete payload

Example:

critic.evaluate(view="technical")
critic.evaluate(view=["executive", "performance"])

---

Session Tracking & Model Comparison

You can persist evaluations and compare model versions over time.

critic_v1 = AICritic(model, X, y, session="v1")
critic_v1.evaluate()

critic_v2 = AICritic(model, X, y, session="v2")
critic_v2.evaluate()

comparison = critic_v2.compare_with("v1")
print(comparison["score_diff"])

This enables:

• Regression tracking
• Risk drift detection
• Governance & audit trails

---

Best Practices & Use Cases

Scenario	Recommended Usage
CI/CD	Block merges using deploy_decision()
Model Tuning	Use technical view for guidance
Governance	Persist session outputs
Stakeholder Reports	Share executive summaries

---

🔒 API Stability

Starting from version 1.0.0, the public API of ai-critic follows semantic versioning. Breaking changes will only occur in major releases.

---

📄 License

Distributed under the MIT License.

---

🧠 Final Note

ai-critic is not a benchmarking tool. It is a decision-making system.

A failed audit does not mean the model is bad — it means the model is not ready to be trusted.

The purpose of ai-critic is to introduce structured skepticism into machine learning workflows — exactly where it belongs.