Irrigation MLOps

ML system for irrigation need prediction (High / Medium / Low) built on FastAPI, Next.js, MLflow, and DVC.

Showcase

Overview — Prediction & Session Stats

Model — Metrics & Feature Importance

Check it on my website, in real deploy Irrigation Demo

Stack

Layer	Tech
API	FastAPI + uvicorn
Frontend	Next.js 16 / React 19 / Tailwind v4
Experiment tracking	MLflow 3.x (basic-auth)
Data versioning	DVC + MinIO S3
Packaging	uv + pyproject.toml
Containers	Docker + Docker Compose

Project Structure

.
├── api/                  # FastAPI application
│   ├── core/             # Config, settings, in-memory job store
│   ├── handlers/         # Route handlers (predict, train, metrics)
│   ├── middleware/        # Rate limiting, API key auth
│   └── schemas/          # Pydantic request/response models
├── pipeline/             # ML pipeline stages
│   ├── preprocess_pipeline.py   # Encode, split, sample weights
│   ├── train_pipeline.py        # Train, log to MLflow, feature importance
│   ├── evaluation_pipeline.py   # Classification report, metrics.json
│   └── predict_pipeline.py      # Inference with MLflow tracing
├── web/                  # Next.js dashboard
│   └── app/
│       ├── components/   # UI primitives and uPlot charts
│       └── page.js       # Main dashboard (Overview / Model / History tabs)
├── data/                 # Raw CSV data tracked by DVC, not git
├── models/               # Trained artifacts — gitignored, mounted as volume
├── notebooks/            # Original competition notebook
├── utils/                # Shared logger
├── main.py               # FastAPI entry point
├── params.yaml           # Single source of truth for all hyperparams
├── dvc.yaml              # DVC pipeline stage definitions
├── Dockerfile            # API image
├── Dockerfile.mlflow     # MLflow image with auth + boto3
└── docker-compose.yml    # Orchestration

Quick Start

# 1. Copy and fill env
cp .env.example .env

# 2. Pull data from MinIO via DVC
dvc pull

# 3. Run locally
uv run uvicorn main:app --reload        # API  → :8000
cd web && npm install && npm run dev    # Web  → :3000
docker compose up mlflow                # MLflow → :5001

API Endpoints

Method	Path	Auth	Description
POST	/api/v1/predict	—	Predict irrigation need (rate limited 20/min)
POST	/api/v1/train	x-api-key	Start background training job
GET	/api/v1/train/{job_id}/status	—	Poll training progress + logs
GET	/api/v1/train/history	—	All training runs this session
GET	/api/v1/metrics	—	Model evaluation metrics
GET	/api/v1/metrics/model-info	—	Current model + MLflow run info
GET	/api/v1/metrics/feature-importance	—	Feature importance scores

Deployment

docker compose up -d --build

Services exposed publicly (joined to alpha_network):

• web → port 3000
• api → port 8000

Services internal only (no host port binding):

• mlflow → accessible only between containers

CI/CD not configured yet. Deployment is currently manual via docker compose on the server. Automated deploy on push to main is planned.

Environment Variables

See .env.example for the full list. Key ones:

MLFLOW_URI                  MLflow tracking server URL
MLFLOW_TRACKING_USERNAME    MLflow basic auth username
MLFLOW_TRACKING_PASSWORD    MLflow basic auth password
MINIO_ENDPOINT              MinIO S3 endpoint
MINIO_ACCESS_KEY
MINIO_SECRET_KEY
API_KEY                     x-api-key required for /train endpoint
NEXT_PUBLIC_API_URL         API URL used by the frontend