Written by a practitioner, for practitioners.

This curriculum exists because I kept having the same conversation with the same kind of engineer, and there was nothing good to hand them afterward.

Who this is for

There is a specific kind of engineer I keep meeting. They shipped real ML — feature pipelines, model training, evaluation, the whole thing in production. Then the LLM stack arrived, moved quickly, and they watched it go past from the outside.

Not because they could not learn it. Because nearly every resource assumes you are either a total beginner or already fluent, and neither describes someone who knows backpropagation cold but has never built a retrieval pipeline or an evaluation harness without ground-truth labels.

This curriculum assumes the first half and teaches the second. It takes linear algebra, training loops, and evaluation intuition as given. It skips what you already know and spends the time on what actually changed.

Where it comes from

I have built and led AI platform work in fintech, healthcare, and telecommunications — industries where innovation has to coexist with regulation. Most recently that has meant architecting a patent-pending agentic risk-assessment system built on LLMs with retrieval-augmented generation and human-in-the-loop review.

Before that, I led science-engineering for Alexa Ranking & Arbitration at Amazon, where my teams built the distributed ML platform behind Alexa's science modeling. Earlier work included AI systems in telecom and healthcare at scale — natural language processing for computer-assisted clinical coding, and platforms handling billions of monthly transactions.

I hold a Ph.D. in Computer Science, with published research in NLP and computer vision. I advise founders, boards, and academic programs, including the Strategic Artificial Intelligence Program at the University of Colorado Colorado Springs.

The last third of this curriculum — evaluation, inference cost, reliability, production patterns — is the part I could not have written fifteen years ago. The gap between a proof of concept and a deployed system is where most of the real engineering lives, and it is the part most courses skip.

How it is built

• Every claim is sourced or marked as opinion. Where I am stating a preference rather than a fact, I say so.
• Every widget documents its math. The calculations are in the footer of each one, so you can check them or adapt them.
• Specifics are dated. Model names and prices change. The engineering frameworks — evaluation discipline, cost arithmetic, reliability stacks — should hold across the next several years. Where something is tied to a particular moment, it says 2026.
• No signup, no upsell, no email gate. Open it in a browser and start.

Licence

Prose is licensed CC BY-SA 4.0. Widget code is MIT. Use it, adapt it, teach from it. Attribution appreciated.

I also write about agentic AI, LLM operationalization, and deploying machine learning in regulated environments at baglodi.com.