mlops-optim

Complete MLOps Pipeline – Training, Serving, CI/CD, DVC, and Webhook Automation

This project implements a full end-to-end MLOps workflow using FastAPI, DVC, Makefile automation, unit tests, API tests, linting, security scans, model quality checks, and an excellence-track webhook with ngrok.

The goal of the project is to build a reproducible ML pipeline for predicting medical insurance costs based on demographic and health factors.

---

1. Project Structure

mlops-optim/
│
├── src/
│   ├── app.py                # FastAPI app (predict / health / model-info)
│   ├── model_pipeline.py     # Prepare → Train → Evaluate pipeline logic
│   ├── prepare_stage.py      # DVC stage: data cleaning + encoding
│   ├── train.py              # DVC stage: model training
│   ├── check_quality.py      # Quality gate: ensure metrics are acceptable
│   ├── webhook.py            # Excellence: remote CI trigger endpoint
│
├── tests/
│   ├── test_api.py           # API tests (health, model-info, predict)
│   ├── test_model_pipeline.py
│
├── data/                     # Raw and intermediate data
├── models/                   # Stored trained models
├── results/                  # Metrics, evaluation results
│
├── dvc.yaml                  # Three-step DVC pipeline
├── dvc.lock
├── Makefile                  # Full CI workflow
├── requirements.txt
└── README.md

This layout follows best practices for modular, maintainable MLOps projects.

---

2. Installation & Setup

1. Create environment

make setup

This creates .make-env/ (Python virtual environment) and installs all dependencies.

2. Activate environment

source .make-env/bin/activate

---

3. Running the DVC Pipeline

To reproduce all ML steps:

make dvc-repro

Stages run automatically:

1. prepare → clean + encode dataset
2. train → GradientBoostingRegressor
3. evaluate → MAE, MSE, R² output to results/metrics.json

Outputs are tracked via DVC.

---

4. FastAPI Model Serving

Start the API:

make run-api

Or manually:

.make-env/bin/uvicorn src.app:app --reload

Available endpoints

Endpoint	Description
GET /health	Service liveness check
GET /model-info	Returns metadata about the loaded model
POST /predict	Returns model prediction for encoded features

Swagger UI

http://127.0.0.1:8000/docs

---

5. How to Make a Prediction

The model expects numerically encoded features:

Feature	Encoding
sex	female=0, male=1
smoker	no=0, yes=1
region	northeast=0, northwest=1, southeast=2, southwest=3

Example input:

{
  "features": [19, 0, 27.9, 0, 1, 3]
}

Example prediction response:

{
  "prediction": 16884.92
}

---

6. Running Tests

Unit tests

make test

API tests

make api-test

All tests must pass before CI succeeds.

---

7. Code Quality, Linting, and Security

Included in CI:

• pylint (score must remain 10/10)
• flake8
• ruff (optional)
• bandit (security scanner)

Run all checks together:

make ci

---

8. CI Pipeline (Makefile)

The Makefile orchestrates the full workflow:

make ci

This runs:

1. pylint
2. flake8
3. bandit
4. unit tests
5. API tests
6. DVC pipeline

If any step fails, CI stops.

---

9. Excellence Track — Webhook CI Trigger

This project includes an advanced MLOps feature:

Webhook endpoint

POST /trigger

Triggers:

make ci

Run webhook server

.make-env/bin/uvicorn src.webhook:app --reload --port 9000

Trigger API

curl -X POST http://127.0.0.1:9000/trigger

---

10. Excellence Track — Ngrok Remote Trigger

Expose the webhook over HTTPS:

ngrok http 9000

You will get a public URL like:

https://cafe-1234.ngrok-free.app/trigger

Trigger CI remotely:

curl -X POST https://cafe-1234.ngrok-free.app/trigger

This demonstrates advanced CI orchestration and earns full excellence marks.

---

11. Model Quality Gate

The check_quality.py script:

• loads results/metrics.json
• verifies required performance thresholds
• stops CI if performance is unacceptable

This prevents merging a degraded model.

---

12. Final Notes

You now have:

• A fully operational ML training pipeline
• A production-like serving API
• A complete CI/CD workflow
• Reproducibility ensured via DVC
• Automated security checks
• A remote CI trigger mechanism
• A fully structured, documented, excellence-level MLOps project

Author : Hind Ch