Model Jacket

Model Jacket is a lightweight toolkit that wraps any ML model (Torch, ONNX, Scikit-learn) into a scalable, observable API server with one command. It provides Dockerized model APIs, metrics, Kubernetes deployment, and a PyPI-installable CLI — all in one place.

---

🚀 Features

• 🐳 One-command containerization with jacket build
• 🔌 Framework-agnostic: PyTorch, ONNX, Scikit-learn supported
• ⚡ FastAPI-based async inference with Uvicorn
• 📊 Built-in Prometheus metrics and /healthz endpoint
• ☸️ Helm chart for scalable K8s deployment + HPA
• ✅ CI/CD to PyPI + GHCR on every Git tag push
• 🧪 Includes batching, versioning, and logging hooks

---

🧑‍💻 Quick Start

1. Train and Export a Model

# export_model.py
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
import joblib

X, y = load_iris(return_X_y=True)
clf = RandomForestClassifier().fit(X, y)
joblib.dump(clf, "model.pkl")

2. Build the Container

jacket build --framework sklearn --model-path model.pkl --tag iris-v1

3. Run Locally

docker run -p 8000:8000 model-jacket:iris-v1

4. Call the API

curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{"input_data": [[5.1, 3.5, 1.4, 0.2]]}'

---

☁️ Deploy with Helm

helm install iris ./helm/model-jacket \
  --set image.repository=ghcr.io/shahwarcodes/model-jacket \
  --set image.tag=iris-v1

---

📦 Install from PyPI

pip install model-jacket

---

📁 Project Structure

model-jacket/
├── jacket_cli/         # CLI for build & serve
├── runtime/            # FastAPI server + handlers
│   ├── api/
│   └── handlers/
├── helm/               # Helm chart for Kubernetes
├── .github/workflows/  # CI/CD pipeline
├── Dockerfile
├── setup.py
└── README.md

---

📈 Observability

• /metrics: Prometheus-compatible metrics (latency, QPS)
• /healthz: Liveness & readiness probes
• Structured logs with model version + timing info

---

🔁 Versioning & Shadow Testing

• Track every model as model:v1, model:v1.1, etc.
• Route production traffic to v1, and shadow to v2
• Compare predictions and latency without user exposure
• Rollback safely using Helm or container tags

---

🤝 Contributing

Contributions welcome! Feel free to fork, submit a PR, or open issues.

---

📄 License

MIT