FoehnCast

FoehnCast tells you which Swiss kiteboarding spot is worth the drive today. It pulls weather forecasts, builds features, trains a quality model, and ranks spots through an API — all following the Feature → Training → Inference (FTI) pattern from the HSLU MLOps course.

Full docs: https://javihslu.github.io/foehncast/

Reviewers: the scoring checklist maps every grading criterion to where it is implemented, with direct links to the code, docs, and live services.

How It Works

flowchart TD
  Forecasts[Weather APIs] -->|Feature pipeline| Curated[Curated features]
  Curated -->|Training pipeline| Registry[MLflow model registry]
  Curated --> Feast[Feast online store]
  Inputs[Live forecast + drive time] --> API[FastAPI]
  Feast --> API
  Registry --> API
  API --> Rank[Ranked spots]

Loading

Three pipelines, one goal:

Pipeline	What it does
Feature	Fetches forecasts, engineers wind features, validates data, stores parquet
Training	Labels quality, trains a model, evaluates it, registers in MLflow
Inference	Serves predictions via FastAPI + ranks spots for one rider profile

Quick Start

Option 1: Try the live demo (no setup)

Open the live demo → — ranked Swiss kiteboarding spots, served from Cloud Run.

What	Link
Streamlit UI	https://foehncast-ui-290885878569.europe-west6.run.app/
API	https://foehncast-serve-290885878569.europe-west6.run.app/
Metrics	https://foehncast-serve-290885878569.europe-west6.run.app/metrics

Option 2: Run locally with Docker

git clone https://github.com/javihslu/foehncast.git
cd foehncast
./scripts/bootstrap-local.sh

The script starts the local stack (Airflow, MLflow, MinIO, Prometheus, and the app) and runs a smoke test. No cloud credentials are needed.

After bootstrap, you get:

Service	URL
App (FastAPI)	http://127.0.0.1:8000
Airflow	http://127.0.0.1:8080
MLflow	http://127.0.0.1:5001
Prometheus	http://127.0.0.1:9090

Try it:

curl -X POST http://127.0.0.1:8000/rank \
  -H 'content-type: application/json' \
  -d '{"spot_ids":["silvaplana","urnersee"]}'

Reproducible pipelines (DVC)

DVC lets you rerun the offline pipelines deterministically:

dvc repro          # ingest → curate → train
dvc metrics show   # check training results

Both DVC stages call the same Python modules as the Airflow DAGs, so the offline and orchestrated runs share one code path.

Tests

make test          # run the unit tests
make lint          # ruff
make coverage      # coverage report

Cloud Deployment

The cloud version runs on GCP Cloud Run. Contributors do not need cloud access; Docker is enough to run everything locally.

flowchart LR
    classDef public fill:#c8e6c9,stroke:#2e7d32,color:#0f2530
    classDef managed fill:#e3f2fd,stroke:#1565c0,color:#0f2530
    classDef svc fill:#fff3e0,stroke:#ef6c00,color:#0f2530

    UI[Streamlit UI]:::public
    APP[FastAPI App]:::public
    MLF[MLflow]:::svc
    WF[Cloud Workflows]:::svc

    BQ[(BigQuery)]:::managed
    GCS[(Cloud Storage)]:::managed
    SQL[(Cloud SQL)]:::managed

    APP --> BQ & GCS
    MLF --> SQL & GCS
    WF -->|schedules pipelines| APP

Loading

Deployment is handled by Terraform (infrastructure) and Cloud Build triggers (images). See the Cloud Architecture docs for details.

Repo Layout

src/foehncast/       # All application code (features, training, inference, monitoring)
dags/                # Airflow DAG definitions
ui/                  # Streamlit dashboard
containers/          # Dockerfiles (6 services)
scripts/             # Bootstrap and helper scripts
terraform/           # GCP infrastructure-as-code
tests/               # pytest unit tests
docs/                # MkDocs source → GitHub Pages
config.yaml          # All tuneable parameters (spots, model, APIs)
dvc.yaml             # Reproducible pipeline stages

Links

• Scoring checklist — grading criteria mapped to evidence
• Full documentation
• Getting started
• Architecture
• Cloud architecture
• Terraform operator detail: terraform/README.md
• Container detail: containers/README.md