Systems and Toolchains for AI Engineers

Systems and Toolchains for AI Engineers#

These are the course notes for Systems and Toolchains for AI Engineers.

Our course is organized around one central idea:

That most of the engineering effort in a deployed artificial intelligence (AI) system lives in the data infrastructure around the model, in the systematic evaluation that tells you whether it works, and in the operational practice that tells you when it has stopped working. This course goes through this whole path with engineering data and case studies in mind.

Where to start#

Read the syllabus for objectives, the toolchain we will be considering in our course, and grading. The schedule maps every session to its topic and to the assignment released or due that week.

The lecture notes are the “textbook” of each session. They are written to be read before or after class, containing extended information and the links to primary sources we considered while preparing this material. The slides used in class are linked from the top of each set of notes.

Every technology, format, method, and case study the course uses is collected in the general index, with each term linked to the section that introduces it. When you remember the idea but not which session covered it, we suggest you start there rather than with search.

Practice modules

Open the practice modules, a short set of questions for each lecture, tied to the numbers and the pitfalls in that session’s notes.

Each module takes about ten minutes, ends by producing a PDF you upload to Canvas, and these will count toward participation. The questions are selected based on your Andrew ID, so the questions you get are not the same as other students.

The same app we created has a “map” of the course, which answers a different question: not “what should I practice” but instead: “which session covered this, and what does it depend on?”. Every corridor on it is a place one lecture actually cites another, so it shows the structure of the course rather than a picture of it. It is not graded, and nothing on it is locked.

Toolchain#

  • Environments: Python, managed with uv.

  • Deep learning: PyTorch and/or JAX.

  • Experiment tracking: MLflow.

  • Data: PostgreSQL, DuckDB, and Parquet.

  • LLMs and agents: framework- and provider-agnostic, since this field changes faster than a semester, in our experience.