CMSC 389A: Modern Software Development with GenAI (UMD, Fall 2026)
by University of Maryland (Anwar Mamat)
Free course notes that take you from Claude Code configuration to RAG, evals and production LLM API calls, with a project in each chapter.
Overview
CMSC 389A is a three-credit University of Maryland undergraduate course taught in Fall 2026 by Senior Lecturer Anwar Mamat, and its public notes ran to 18 chapters as of November 3, 2026. Chapters 1 to 12 cover Claude Code and full-stack basics: CLAUDE.md, settings scopes, subagents, skills and hooks, then Git and CI/CD, FastAPI, MySQL, React with TypeScript, authentication, layered architecture, Docker and Redis. Chapters 13 to 18 turn to LLMs. The LLM chapter follows Andrej Karpathy's lecture "Deep Dive into LLMs like ChatGPT" and adds TerpLM, a miniature model you build, sample from and post-train on a laptop. Chapter 14 teaches the Elephant-Goldfish Model for writing and testing design documents. The RAG chapter builds chunking, BGE-M3 embeddings, a vector index, BM25 and Reciprocal Rank Fusion, then exposes retrieval as a Claude tool. Later chapters cover evals with promptfoo, a six-service local RAG stack using Ollama, Infinity and ChromaDB, and production API work with streaming, retries, cost limits and LiteLLM. Enrolled students pay $20 a month for Claude Code Pro; reading the notes costs nothing. The notes index shows no license, some companion repos sit on UMD's GitLab, and the schedule listed no topics after October 29 when checked. A May 2026 UMD article says Mamat and Assistant Professor Leonidas Lampropoulos received a grant to develop an undergraduate course called Modern Software Development.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- 16-week university semester (Aug 31 to Dec 18, 2026); 18 public chapters of notes, self-paced
- Provider
- University of Maryland (Anwar Mamat)
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Write allow, ask and deny permission rules and predict which Claude Code settings scope wins
- ✓Create a custom slash command, a subagent and a hook, then explain who triggers each
- ✓Build, sample from and post-train a miniature language model on a laptop with TerpLM
- ✓Run the Goldfish Protocol's comprehension, critic and implementation-readiness tests on a design document
- ✓Build a multi-index retriever that combines BM25 and embeddings using Reciprocal Rank Fusion
- ✓Write code-graded and model-graded evals by hand in Python and declaratively with promptfoo
- ✓Orchestrate a six-service local RAG stack with Docker Compose, Ollama, Infinity and ChromaDB
- ✓Implement exponential backoff with jitter and compute a request's dollar cost from token usage
Highlights
- •Agent configuration, web engineering and LLM internals sit in one sequence, so RAG and evals get built inside the same kind of application stack you already ship
- •The RAG mini-project uses the standard library for chunking and BM25, and its tests run offline with fake embeddings
- •The local LLM chapter runs on a recent laptop CPU using qwen2.5:1.5b, about 1 GB on disk, and names the bugs you will hit, including silent context truncation and mis-ordered embeddings
- •The chapters checked end in practice exercises, and several also list common mistakes
- •Coursicle and PlanetTerp list the course but carried no reviews specific to it on November 3, 2026, so there is no outside report yet on its difficulty
Who It’s For
Best For
- ✓Backend or full-stack developers adding RAG, evals and LLM API calls to the applications they already build
- ✓Developers using Claude Code who want a structured treatment of its configuration files and a design-first workflow
- ✓Instructors designing a university course that teaches software engineering with coding agents
Prerequisites
- •UMD requires a C or better in CMSC 216 and CMSC 250, so expect to need solid programming fundamentals
- •A Claude Code Pro subscription ($20/month) to follow the agent chapters as enrolled students do
- •Docker and Docker Compose for the full local RAG stack in the Ollama chapter
FAQ
What is CMSC 389A: Modern Software Development with GenAI (UMD, Fall 2026)?
CMSC 389A is a University of Maryland undergraduate course whose public notes teach software development with coding agents, for programmers who want agent workflows and LLM engineering in a single sequence. After working through it, you can configure Claude Code for a team, build a hybrid RAG pipeline, write code-graded and model-graded evals, and run a local LLM stack.
Is CMSC 389A: Modern Software Development with GenAI (UMD, Fall 2026) free?
CMSC 389A: Modern Software Development with GenAI (UMD, Fall 2026) is free to access.
What level is CMSC 389A: Modern Software Development with GenAI (UMD, Fall 2026) for?
CMSC 389A: Modern Software Development with GenAI (UMD, Fall 2026) is aimed at a intermediate audience. Recommended background: UMD requires a C or better in CMSC 216 and CMSC 250, so expect to need solid programming fundamentals, A Claude Code Pro subscription ($20/month) to follow the agent chapters as enrolled students do, Docker and Docker Compose for the full local RAG stack in the Ollama chapter.
How long does CMSC 389A: Modern Software Development with GenAI (UMD, Fall 2026) take?
Expect roughly 16-week university semester (Aug 31 to Dec 18, 2026); 18 public chapters of notes, self-paced. Most learners work through it at their own pace.
What will I learn from CMSC 389A: Modern Software Development with GenAI (UMD, Fall 2026)?
You'll learn: Write allow, ask and deny permission rules and predict which Claude Code settings scope wins; Create a custom slash command, a subagent and a hook, then explain who triggers each; Build, sample from and post-train a miniature language model on a laptop with TerpLM; Run the Goldfish Protocol's comprehension, critic and implementation-readiness tests on a design document; Build a multi-index retriever that combines BM25 and embeddings using Reciprocal Rank Fusion; Write code-graded and model-graded evals by hand in Python and declaratively with promptfoo; Orchestrate a six-service local RAG stack with Docker Compose, Ollama, Infinity and ChromaDB; Implement exponential backoff with jitter and compute a request's dollar cost from token usage.
Topics
Sources
This page was written from 7 sources, 2 on domains other than cs.umd.edu.