All 60 tasks, level by level

The real graded MLOps tasks the portal assigns, grouped into three levels.

The portal assigns these MLOps tasks. Finish Level 1 before Level 2 and Level 2 before Level 3. Your mentor may mark some tasks as required for the certificate.

Level 1: Foundations (24 tasks)

Task	What to do
What MLOps is	Explain the ML lifecycle (data, training, deployment, monitoring) and where MLOps fits, with a diagram.
Reproducible environment	Create a project with a pinned requirements file or conda environment and a README that anyone can follow.
Train a baseline model	Train a simple scikit-learn model on a public dataset and record its accuracy and training time.
Project structure	Organise your ML code into data, features, train and predict modules instead of one notebook.
Version code with Git	Commit your project with a clear history and a .gitignore that keeps data and models out of Git.
Version data with DVC	Track your dataset with DVC and show how to switch between two versions of the data.
Experiment tracking	Log parameters, metrics and the model for 5 training runs with MLflow and compare them.
Save and load models	Save a trained model with joblib or ONNX and load it in a separate script to predict.
Model serving API	Serve your model with FastAPI or Flask: POST features, get a prediction back as JSON.
Input validation	Validate prediction requests with Pydantic and return clear errors for bad input.
Unit tests for ML code	Write tests for data cleaning, feature code and the prediction endpoint.
Containerise the model	Write a Dockerfile for the model API and run it locally.
Makefile or task runner	Add make (or invoke/just) commands for setup, train, test and serve.
Config files for training	Move hyperparameters to a YAML config and load it with Hydra or a simple loader.
Random seeds	Make training reproducible with fixed seeds and show two identical runs.
Data schema	Define the expected input schema with Pandera or Great Expectations and validate it.
Model card	Write a model card: purpose, data, metrics, limits and ethical considerations.
Notebook to script	Convert an exploratory notebook into clean, tested Python modules.
Pre-commit for ML	Add black, ruff and nbstripout as pre-commit hooks.
Environment with Docker	Build a training image so anyone can run training with one command.
Compare model formats	Save a model as pickle, joblib and ONNX and compare size and load time.
Simple dashboard of runs	Show your MLflow runs in a table and pick the best model with a rule.
API contract	Document your prediction API with OpenAPI, including examples.
Cost of training	Estimate the cost of training your model on a cloud GPU vs a CPU.

Level 2: Building real features (20 tasks)

Task	What to do
Training pipeline	Turn data preparation, training and evaluation into a pipeline (DVC pipeline, Prefect or Airflow).
CI for ML	Add GitHub Actions that run tests and a quick training check on every push.
Model registry	Register models in MLflow's registry and move one from staging to production.
Deploy to the cloud	Deploy the model API to a free cloud service and share the live URL.
Batch prediction	Build a batch job that scores a whole CSV file and writes the results with a timestamp.
Feature store concept	Explain feature stores and build a simple shared feature table used by training and serving.
Hyperparameter tuning	Tune your model with Optuna or GridSearch and log every trial to MLflow.
Model evaluation report	Produce an evaluation report with metrics, confusion matrix and error analysis on slices of data.
Load testing	Load-test your model API with Locust or k6 and report latency and throughput.
Logging predictions	Log every prediction with inputs, output, model version and time for later monitoring.
Feature pipeline	Separate feature computation into a reusable pipeline used by training and serving.
Pipeline orchestration	Run your pipeline in Airflow, Prefect or Kubeflow Pipelines with retries.
Model validation gate	Block promotion of a model that performs worse than the current one on a holdout set.
Shadow deployment	Run a new model in shadow mode next to the old one and compare predictions.
Serving with BentoML or Ray Serve	Serve your model with a model-serving framework and compare with FastAPI.
GPU basics	Train a small deep learning model on a free GPU notebook and log GPU usage.
Prediction caching	Cache predictions for repeated inputs and measure latency improvement.
Secrets and credentials	Keep cloud and database credentials out of code in your ML pipeline.
Data versioning workflow	Show a full workflow: change data, retrain, compare and roll back with DVC.
Explainability	Add SHAP explanations to predictions and show them in the API response.

Level 3: Professional work and capstone (16 tasks)

Task	What to do
Data drift detection	Use Evidently or a custom check to compare new data with training data and flag drift.
Model monitoring dashboard	Build a dashboard with prediction volume, latency and drift over time.
Automated retraining	Trigger retraining when drift is detected or on a schedule, and register the new model only if it beats the old one.
Kubernetes deployment	Deploy the model to a local Kubernetes cluster with a health check and two replicas.
`A/B or canary release`	Route a share of traffic to a new model version and compare their results.
LLM serving basics	Serve a small open model or an LLM API behind your own endpoint with caching and rate limits.
Responsible ML checks	Check your model for bias across one sensitive attribute and document the result and fix.
Capstone: end-to-end MLOps	Deliver a full MLOps project: versioned data, tracked experiments, CI, deployed API, monitoring and retraining, with an architecture diagram.
Concept drift simulation	Simulate concept drift on a dataset and show your monitoring catching it.
Feature store with Feast	Set up Feast with offline and online stores for one feature set.
LLMOps basics	Version prompts, log LLM calls and evaluate outputs for an LLM feature.
RAG evaluation pipeline	Build an automated evaluation for a RAG app with a fixed question set.
Autoscaling model service	Configure autoscaling for your model service and test it under load.
Governance and audit	Record who trained, approved and deployed each model version.
Incident runbook for models	Write a runbook for a model producing bad predictions in production.
Capstone 2: ML platform template	Build a reusable MLOps template repo (data checks, training, tracking, CI, serving, monitoring) and document how a team would use it.

← Tools and first-week setup Submitting tasks and earning your certificate →