AI/ML Systems for Executives

What you build

One system, assembled over ten weeks: an AI analyst that reads a company's documents — a board deck, an investor memo, a customer contract — and answers questions about them with cited sources.

By Week 9 your system will:

• —Ingest PDF documents, including scanned pages, using text extraction and OCR
• —Split documents into indexed passages (chunking) and convert them into a searchable semantic index (embeddings, vector search)
• —Retrieve the relevant passages for any question and generate an answer with citations (retrieval-augmented generation)
• —Be scored against a 20-question evaluation set you wrote, with known answers
• —Carry a guardrail you authored — a written, testable rule the system may not break
• —Use one external tool on its own (an AI agent, governed)
• —Have a cost model: what it costs today, and what it costs at 10,000 documents

You direct the build; an AI assistant does the typing. Each week ships the exact prompts to use.

The tools — and what each one is for

Everything below is free. The stack is not the curriculum — each tool exists to produce something you keep.

Google Colab

The workbench. Each week's build runs in your browser; nothing to install, nothing to break.

What you keep: A working system you can rerun, extend, or show anytime.

Claude

The builder. You direct it with provided prompts; it writes the code. This is how most software gets built now — directing AI rather than typing — and the coding involvement here is exactly that, at most.

What you keep: The skill of directing AI work and judging the result, which this course pairs with the verification discipline that practice usually lacks.

Groq

The engine. Runs the AI model behind your system, with a live per-token price meter.

What you keep: A real cost model, at today's usage and at 10,000 documents.

Databricks

The enterprise view. One session inside the data platform large companies actually run (Week 8).

What you keep: Recognition — you'll know what your company is buying and what the invoice pays for.

Google Sheets

The scorecard. Your 20-question evaluation set lives in a spreadsheet.

What you keep: The scoring artifact to request from any vendor.

GitHub

The record. The course materials live here; optionally, so does your finished build.

What you keep: Proof of work.

Who this is for

Senior executives — CIOs, CTOs, COOs, GMs, board members — whose technical careers were built before AI. You've run systems and led engineering organizations. Models, embeddings, and inference arrived after your hands-on years. This course puts your hands on the technology at the right depth: enough to build it, test it, question it, and speak its working vocabulary naturally — in the boardroom and with your engineers.

What you learn

1. 01

   The architecture.

   AI systems are systems — ingest, index, retrieve, generate, monitor. You'll know each component because you assembled it.

2. 02

   The vocabulary.

   Embeddings, RAG, chunking, inference, agents, guardrails, hallucination — used correctly and naturally, because each term names something you built.

3. 03

   The verification discipline.

   AI output is fluent and confident even when it's wrong, and confirmation bias makes fluent answers hard to doubt. You'll test your own system, find its wrong answers, categorize them, and score it against ground truth. We anticipate the known issues — retrieval misses, confident invention, stale data, silent data errors — and you resolve each one in your own system, so you recognize them anywhere.

The three phases

How it ends

In Week 9 you document your system on one page — what it does, what it costs at current and 10x scale, its evaluation score, its three known limitations, and its guardrail. This System Card is the disclosure to request from any vendor, written first about your own system. Submitting it issues your Instructor Endorsement — a statement of verifiable work, not attendance.

What this course will not teach you

Five topics come up constantly in AI conversations, and the course deliberately leaves each one closed. Here is what is behind each door, and why you will not need to open it.

Gradient descent, and local versus global minima

How a model finds good weights during training: nudge every number in the direction that reduces error, billions of times over, while risking settling in a low spot that is not the lowest.

You will never train a model. You will rent one a lab spent nine figures training, so the training procedure is not a decision you get to make.

Backpropagation

The bookkeeping that makes gradient descent computable: each error is assigned backward through the network so every weight knows its share of the blame.

It finished in a data center before you typed your first prompt. It is frozen in the weights, and nothing you do at runtime changes it.

Fine-tuning mechanics

Continuing a model's training on your own data to specialize it.

You need two facts about it, not the procedure: it exists, and it is usually the wrong first answer. Retrieval plus a well-written system prompt beats fine-tuning for most corporate use cases at a fraction of the cost and risk. The useful output is a question — why fine-tuning instead of retrieval?

Transformer internals beyond attention

Layer normalization, residual streams, positional encodings — the plumbing that holds a transformer together.

You learn attention in Week 3, because it explains why your system repeated a false fact you planted. The rest explains nothing you will ever observe in your own system.

CUDA and GPU programming

The low-level programming model that makes graphics hardware run model math efficiently.

You will price GPU inference in Week 8. You will never program one.

Questions

Do I need to know how to code?

No. You direct an AI assistant using provided prompts; it writes the code. Reading it is optional.

What does it cost?

Nothing, permanently. All ten weeks, the builds, and the endorsement are free. A paid tier may exist later (personal capstone review, live groups); the course itself stays free.

What's the time commitment?

The material is organized as ten weekly modules of about two hours each. It's fully self-paced: a steady two hours a week finishes in ten weeks; a focused quiet weekend gets you through the first phase or further. Join any time, progress on your own schedule.

What do I need?

A browser and three free accounts, all created in Week 0.

Whose documents do I use?

The course ships with a complete fictional company: a board deck, an investor memo, and a customer contract, built realistically for the exercises. You never need to upload anything of your own. If you choose to use your own documents, apply one test first: would you share them with an outside vendor?

I'm more technical than your typical student.

Then Weeks 0–2 will be fast for you and Weeks 3–9 — evaluation, silent failures, cost modeling — are the part most technical people have also never done.

About the instructor

Bill Vallier has spent thirty years building enterprise data systems, operating as both an executive and a working engineer with code in production today. The course's examples come from real systems, not textbooks. This is its first run; there are no testimonials yet, and the course is free partly for that reason — the first students are the proof.

Begin at Week 0

Rolling enrollment. Two hours a week, at your own pace.