Pipeline
Train-to-serve
Pinned data, registry, a gated deploy, a rollback switch.
MLOps Deep Dive
Training is not the job. Pipelines, registries, promotion, monitoring, and a rollback when the new model is worse.
Created by Baljeet Dogra
Course Objectives
- Promote a model because of evals, not because the loss looked nice.
- Reproduce a run from a commit and a data pin.
- Watch live quality, not only latency.
- Roll back in a drill, not in a post-mortem.
What you will produce
Ops
Quality dashboard
Online proxy metrics and an alert you have actually fired.
6 weeks curriculum
Expand a part for the syllabus. Content stays searchable when closed.
01 Reproducibility Week 1
Seeds, data pins, environments, experiment tracking without theatre.
- Pins
- Tracking
- Seeds
- Artefacts
02 Registry and CI Weeks 2–3
Model cards, promotion rules, evals that can fail the build.
- Registry
- Cards
- Gates
- CI
03 Deploy patterns Week 4
Batch, online, shadow, canary. Feature freshness.
- Batch
- Online
- Shadow
- Canary
04 Monitor and roll back Weeks 5–6
Drift, data quality, quality proxies. Capstone drill.
- Drift
- Alerts
- Rollback
- Capstone
Who this is for
ML engineers in production
You train. Ops still happens in a notebook and a hope.
Platform engineers
You need ML-shaped CI, not a copy of the app pipeline.
Prerequisites
- You have trained a model
- Git and CI at a basic level
- Containers at a basic level
Not a fit if
- Kubeflow tourism
- People who have never trained anything
Related: How to Become an AI Engineer · Kubernetes for AI Systems
Questions
Which stack?
One tracking tool, one registry, one deployer. Vendor-plural on purpose.
Is this Kubernetes?
Optional. Kubernetes for AI Systems is the dedicated course.
Ready to start?
Six weeks to promote, watch and roll back a model. Create an account to enrol.
Enrol now