MLOps Mastery Path¶

Mastery Path

BOOTCAMP

VINOD BHANUJA

Most machine learning tutorials stop the moment a model scores well in a notebook — and skip the much harder, much more common real problem: six months later, nobody can remember which data trained the model currently in production, a retrain quietly made things worse, and nobody noticed until a customer did. MLOps is the discipline that closes that gap, and this course builds every piece of it for real: real experiment tracking against a real local server, a real model registry with real version promotion, a real CI/CD pipeline that actually blocks a bad model from shipping, and real drift detection computed from real, deliberately-shifted data.

This course was written and run against a real local MLflow tracking server, real scikit-learn models trained on real data, and real deployment pipelines — no cloud ML platform account, no managed service, no synthetic screenshots standing in for a UI nobody actually clicked through. Real training runs, logged for real; a real model promoted from staging to production and actually served; real drift, injected and actually detected.

It assumes basic Python and command-line comfort, and some exposure to the model-building side of machine learning — this course is about what happens around a trained model, not how to train one from scratch. For the modeling fundamentals themselves, see the AI/ML Mastery Path; this course deliberately doesn't repeat that material.

How the program is organized¶

Level	Focus	Modules	What you can do after
1 · Experiment Tracking & Reproducibility	Runs, params, metrics, artifacts, hyperparameter sweeps, reproducibility	10	Track real experiments so any result can be reproduced and compared
2 · Model Registry & Packaging	Versioning models, stages, signatures, approval workflows, real local serving	10	Register, promote, and serve a real model version with confidence
3 · Pipelines & Deployment	Training pipelines, CI/CD for ML, containerizing, canary/shadow deployment, feature stores	10	Ship a model through a real, automated, tested pipeline
4 · Monitoring, Drift & Operating ML at Scale	Data/concept drift, monitoring without ground truth, retraining triggers, governance	10	Detect real model degradation and decide when to retrain

Each level has 9 teaching modules plus a hands-on project (module 10), for 40 lessons total.

How to use this site¶

Run it yourself. A real local MLflow server (pip install mlflow) and scikit-learn are all this course's core lessons need — no account, no cloud bill, no managed ML platform.
Read the "what was run" note at the top of each lesson. The overwhelming majority of this course's metrics, comparisons, and detected drift come from real, observed runs; anything that genuinely couldn't be run locally (a full managed feature store, a large-scale distributed training job) says so plainly.
Do the "predict, then run" exercises. MLOps has several behaviors (training/serving skew, drift that doesn't show up in accuracy alone, autologging gaps) that are easy to miss from reading alone and genuinely clarifying to watch happen.
Do the projects. Each one is a real, runnable tracking setup, registry workflow, or pipeline, not a diagram.