ML Project Template for 2025 - Build ML Pipelines with Python, uv, DVC, FastAPI, Docker

A ready-to-use machine learning project template is positioned as a 2025 blueprint for taking a model from dataset creation to a production-style REST API and a Dockerized deployment. The core payoff is a structured pipeline that combines reproducible data and training steps (via DVC), a FastAPI service layer for inference, and a Docker workflow that packages the whole system so it can run on a VPS or other hosting environment.

Setup starts with cloning a public GitHub repository and using uv to manage Python and dependencies. The workflow creates a virtual environment pinned to a specified Python version, installs project dependencies quickly, and installs the project itself in editable mode so local code changes are importable during development. A pre-commit hook is then installed to enforce formatting and checks before changes land in Git—using tools like Ruff for formatting/linting and pytest for testing.

The template’s structure is organized around standard ML engineering directories: artifacts for outputs, bin for scripts, a profit package for the application code, tests for verification, plus Dockerfile and app.py for the serving layer. Dependencies listed include DVC for data versioning, FastAPI and Uvicorn for the API server, pandas for data handling, and FastAPI-related tooling (such as fastparquet for Parquet I/O). The template also includes configuration and tooling for development hygiene (pre-commit, Ruff, pytest).

At the heart of the pipeline is a REST API endpoint that serves predictions from a text classifier. The API reads a config file, loads or references the model used for inference, and exposes a single predict endpoint. A client sends text via POST, and the service returns a JSON response containing the predicted label. In the example shown, the classifier returns “just do it” for the provided input text.

Reproducibility and workflow automation are handled through a DVC pipeline defined in a dvc.yaml file. The pipeline uses three common stages: build the dataset, train the model, and evaluate the model. The dataset stage runs a bash script that seeds and configures the environment, then calls a dataset builder that constructs labeled examples (e.g., mapping text to labels like “just do it” versus “stop doing it”). The builder splits data into training and testing subsets using pandas operations and writes them out as Parquet files (train.parquet and test.parquet) according to paths specified in the config. DVC then tracks these artifacts and enables the steps to be reproduced consistently.

Training and evaluation are run as subsequent DVC stages. The example uses a dummy model but still produces an evaluation metric—reported as accuracy of 1—demonstrating the end-to-end flow from data artifacts to measurable performance.

Finally, deployment readiness comes from Docker. A Dockerfile is included so the image can be built and run locally; when started, the container exposes the FastAPI documentation and the same predict endpoint on the expected local host URL. Overall, the template is presented as a fast way to bootstrap an ML pipeline that is reproducible, testable, serves inference via HTTP, and packages cleanly for deployment.