driftguard-ai-sdk 1.0.2

Production-grade AI model monitoring, drift detection, and autonomous retraining platform

DriftGuard — Autonomous Model Health Platform

DriftGuard is a production-grade, self-healing MLOps platform designed to detect data drift, concept drift, and model degradation in real time, automatically trigger validating retraining pipelines, and progressively deploy champion models via progressive canary routers.

---

Architecture Design

                     +---------------------------------------+
                     |          Client Application           |
                     +-------------------+-------------------+
                                         |
                                (Predict Telemetry)
                                         v
                     +-------------------+-------------------+
                     |          DriftGuard SDK               |
                     |  - Wrapper pattern intercept          |
                     |  - River ADWIN concept drift checks   |
                     +-------------------+-------------------+
                                         |
                                 (HTTP Telemetry)
                                         v
                     +-------------------+-------------------+
                     |       DriftGuard FastAPI Core API     | <---+ NextJS Dashboard (:3000)
                     |       - /register, /predict, /drift   | <---+ Grafana (:3001)
                     |       - Prom metrics /metrics (:8000) |
                     +-------------------+-------------------+
                                         |
                        (SLA Drift Breach Trigger)
                                         v
                     +-------------------+-------------------+
                     |      Prefect Orchestration Server     |
                     |      - drift_detection_flow (:4200)   |
                     +-------------------+-------------------+
                                         |
                                 (Runs steps)
                                         v
                     +-------------------+-------------------+
                     |      ZenML Step Training Pipelines    |
                     |  - Step 1: Great Expectations Validate|
                     |  - Step 2: Feast Feature Store Check  |
                     |  - Step 3: Train & Track (MLflow/W&B) |
                     |  - Step 4: Validate (>1% boost check) |
                     |  - Step 5: Canary Progressive Deploy  |
                     |  - Step 6: Immutable JSON Ledger & PDF|
                     +-------------------+-------------------+
                                         |
                            (Progressive Split Promotes)
                                         v
                     +-------------------+-------------------+
                     |       BentoML & Ray Serve Fleet       |
                     |       - canary_router: 10%->100%      |
                     |       - SLA Monitoring & Rollbacks    |
                     +---------------------------------------+

---

Prerequisites

To run and configure DriftGuard, ensure the following are installed:

• Python 3.11 only (Ray and BentoML have incomplete 3.12 support).
• Docker & Docker Compose (for multi-service orchestration).
• kubectl & Helm (optional, for Kubernetes deployments).
• HashiCorp Terraform (optional, for AWS cloud provisioning).

---

Quick Start in 5 Lines

Wrap any scikit-learn, PyTorch, or HuggingFace model with DriftGuard SDK to track predictions, compute concept drift, and initiate auto-healing:

from driftguard import DriftGuard

# 1. Initialize DriftGuard
dg = DriftGuard(model_id="fraud-detector-v1", api_url="http://localhost:8000", drift_threshold=0.15, auto_retrain=True)

# 2. Wrap model seamlessly
model = dg.wrap(trained_sklearn_model)

# 3. Predict normally - DriftGuard tracks inputs, outputs, and triggers retrain on drift!
prediction = model.predict(features)

---

Installation & Setup

1. Local Package Installation

Clone the repository and install the DriftGuard package locally:

git clone https://github.com/your-repo/DriftGuard.git
cd DriftGuard
pip install -e .

To install validation pipelines dependencies (Great Expectations + SQLAlchemy 1.4 pin) separately:

pip install -e ".[validation]"

2. Launch Platform Services

Spin up the entire 8-service DriftGuard stack (FastAPI, NextJS dashboard, MLflow, Prefect, Postgres, Redis, Prometheus, Grafana) instantly:

docker-compose -f infra/docker-compose.yml up --build -d

---

SDK Configuration Parameters

The DriftGuard class accepts the following parameters:

Parameter	Type	Default	Description
model_id	str	Required	Unique name identifier of the model.
api_url	str	http://localhost:8000	Gateway endpoint of the DriftGuard API.
drift_threshold	float	0.15	Limit before concept drift alert and retraining triggers.
auto_retrain	bool	True	Automatically triggers retraining flow on threshold breach.

---

API Gateway Documentation

DriftGuard Core API runs on port 8000. Key REST endpoints include:

POST /register

Registers a model for monitoring.

• Request Body:

  {
    "model_id": "fraud-detector-v1",
    "drift_threshold": 0.15,
    "reference_data_path": "./data/baseline.parquet",
    "features": ["amount", "location_score", "velocity"]
  }
  

• Response: {"status": "registered", "model_id": "fraud-detector-v1"}

POST /predict/{model_id}

SDK telemetry gateway recording prediction details and updating gauges.

• Request Body:

  {
    "features": [1.2, 0.4, 9.8],
    "prediction": [1.0],
    "drift_score": 0.08
  }
  

GET /drift/{model_id}

Fetches the last 100 historical drift scores for Recharts charts rendering.

POST /retrain/{model_id}

Manually triggers the background retraining pipeline flow.

GET /metrics

Exposes system health gauges for Prometheus scrapers in OpenMetrics format.

---

Dashboards & Observability Portals

Once the docker services are online:

1. NextJS UI Dashboard: Navigate to http://localhost:3000 to review models list, drift histories, vertical retraining timelines, and searchable audit logs.
2. MLflow Registry: Access http://localhost:5000 to review runs parameters, artifacts (confusion matrix plots), and staging/production champions.
3. Prefect Dashboard: Access http://localhost:4200 to inspect flows execution history.
4. Grafana Dashboards: Open http://localhost:3001 (User: admin | Pass: admin) to view pre-provisioned telemetry panels scraping predictions, drift rates, accuracy levels, and quantiles latency.

---

Running Unit Tests

Run all unit tests verifying ADWIN concept detectors, Great Expectations validators, canary routing splits, emergency rollbacks, and cryptographic audit log chains:

pytest tests/ -v

---

Deploying to AWS Cloud (Terraform)

Deploy DriftGuard core infrastructure to Amazon Web Services:

cd infra/terraform
terraform init
terraform plan
terraform apply -var="db_password=SecurePasswordPass22!"

This provisions:

• Amazon EKS cluster for progressive Kubernetes canary Rollouts.
• Amazon RDS PostgreSQL database for MLflow, Prefect, and platform metadata.
• Amazon ElastiCache Redis for online real-time Feast features access.
• Amazon S3 bucket for artifacts.
• Amazon ECR for Docker images.