Machine Learning in Production / AI Engineering (CMU 17-645, 11-695, Spring 2026)
by Carnegie Mellon University
The course on everything around the model — requirements, deployment, monitoring, security, fairness — with its MIT Press textbook free online.
Overview
The Spring 2026 schedule runs 24 lectures from January to April: Introduction and Motivation; Correctness and Risk; Setting Goals and Gathering Requirements; Planning for Mistakes; Model Quality; Fostering Interdisciplinary Teams; Behavioral Model Testing; Automating and Testing ML Pipelines; Deploying a Model; Testing and Experimenting in Production; Data Quality; Scaling the System; Planning for Operations; ML Security; System Security; Safety; Process and Technical Debt; Versioning, Provenance and Reproducibility; Explainability; Transparency; Accountability and Ethics; Measuring Fairness; and Building Fairer Systems. Coursework is four individual assignments (I1 ML Product, I2 Requirements, I3 MCP and Security, I4 Explainability), a four-milestone group project covering modeling and first deployment, infrastructure quality, monitoring and continuous delivery, then fairness, security and feedback loops, plus 11 labs on Flask APIs, Kafka, Git, model testing, Docker, CI/CD, agents and MCP, monitoring, provenance tooling, explainability tooling and Minikube. The course is 12 units, graded 35% individual assignments, 30% group project, 25% midterms and participation, 10% labs. The companion textbook, Machine Learning in Production: From Models to Products (MIT Press, April 2025), is open access under a Creative Commons licence; slides are published publicly, all course material sits in the mlip-cmu GitHub organisation (the s2026 repository was updated in August 2026), and lectures are recorded and released on YouTube. Christian Kaestner is an associate professor in CMU's Institute for Software Research, director of its Software Engineering PhD program, and the textbook's author.
At a Glance
- Topic
- ML
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- 15 weeks (Jan-Apr 2026): 24 lectures, 11 labs, 4 assignments and a group project; ~12h/week at full 12-unit load
- Provider
- Carnegie Mellon University
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Translate a product goal into ML requirements and measurable system-level quality targets
- ✓Design for inevitable model mistakes using guardrails, fallbacks and human oversight
- ✓Build and test automated ML pipelines instead of one-off notebook training runs
- ✓Deploy a model behind an API and run online experiments safely in production
- ✓Instrument monitoring and continuous delivery so silent model degradation is caught
- ✓Threat-model ML and agent systems, including a dedicated MCP and security assignment
- ✓Track versioning, provenance and reproducibility across data, models and pipelines
- ✓Measure fairness and apply explainability tooling to an already-deployed system
Highlights
- •The companion MIT Press textbook is open access under Creative Commons — a full published book, not lecture notes
- •Every artifact is public: slides, 11 labs, four assignments, the group project spec, GitHub repos, and lectures recorded to YouTube
- •Covers the work modeling courses skip entirely — requirements, operations, security, provenance, technical debt and fairness
- •The Spring 2026 revision adds an agents-and-MCP lab and an 'MCP and Security' assignment, so it tracks how agents are actually deployed now
- •Taught by the textbook's author, with a self-assessment prerequisite quiz so you can check your readiness before committing
Who It’s For
Best For
- ✓ML engineers whose models work in a notebook but keep breaking in production
- ✓Software engineers newly responsible for shipping an AI feature end to end
- ✓Tech leads and MLOps practitioners who need shared vocabulary for AI system quality
- ✓Self-learners who want a complete university course with a free published textbook
Prerequisites
- •Basic fluency in Python, plus comfort with a Unix shell and installing libraries
- •Familiarity with feature extraction, model training and evaluation, and LLM prompting (a self-check quiz is provided)
- •No prior software engineering coursework required — requirements, design and testing are taught from scratch
FAQ
What is Machine Learning in Production / AI Engineering (CMU 17-645, 11-695, Spring 2026)?
Carnegie Mellon's Spring 2026 offering of 17-645 Machine Learning in Production, cross-listed as 11-695 AI Engineering, taught by Christian Kaestner and Claire Le Goues. It teaches the software engineering around an ML or LLM component — requirements, deployment, pipeline testing, monitoring, security, provenance and fairness — rather than modeling itself. Slides, labs, assignments and the MIT Press textbook are all public, so a self-learner can follow the entire course for free.
Is Machine Learning in Production / AI Engineering (CMU 17-645, 11-695, Spring 2026) free?
Machine Learning in Production / AI Engineering (CMU 17-645, 11-695, Spring 2026) is free to access.
What level is Machine Learning in Production / AI Engineering (CMU 17-645, 11-695, Spring 2026) for?
Machine Learning in Production / AI Engineering (CMU 17-645, 11-695, Spring 2026) is aimed at a intermediate audience. Recommended background: Basic fluency in Python, plus comfort with a Unix shell and installing libraries, Familiarity with feature extraction, model training and evaluation, and LLM prompting (a self-check quiz is provided), No prior software engineering coursework required — requirements, design and testing are taught from scratch.
How long does Machine Learning in Production / AI Engineering (CMU 17-645, 11-695, Spring 2026) take?
Expect roughly 15 weeks (Jan-Apr 2026): 24 lectures, 11 labs, 4 assignments and a group project; ~12h/week at full 12-unit load. Most learners work through it at their own pace.
What will I learn from Machine Learning in Production / AI Engineering (CMU 17-645, 11-695, Spring 2026)?
You'll learn: Translate a product goal into ML requirements and measurable system-level quality targets; Design for inevitable model mistakes using guardrails, fallbacks and human oversight; Build and test automated ML pipelines instead of one-off notebook training runs; Deploy a model behind an API and run online experiments safely in production; Instrument monitoring and continuous delivery so silent model degradation is caught; Threat-model ML and agent systems, including a dedicated MCP and security assignment; Track versioning, provenance and reproducibility across data, models and pipelines; Measure fairness and apply explainability tooling to an already-deployed system.
Topics
Sources
This page was written from 3 sources, 2 on domains other than mlip-cmu.github.io.