Privacy-First AI Systems Training

Why We Teach

AI changes what organizations can do with their data.

For decades, we built data systems primarily to organize structured information for reporting and analytics. But much of an organization's knowledge lives outside those systems — in documents, notes, reports, messages, images, and free text.

Modern AI lets us work with that information at a scale and ease we haven't seen before. It opens new ways to work with our data, and new ways to make that data valuable. But a model alone isn't enough — it needs the right context, the right data, and systems that can retrieve, assemble, evaluate, and protect that information.

That creates a new job for data teams — and a new set of skills they need to learn.

How We Teach

Two ideas run through every module: building the systems — agentic ones included — that put AI to work, and choosing the right models to put inside them.

Systems

A model is only one part of a working AI application. Systems connect models to people, data, and other tools. They manage what goes into a model and what happens to its output, connect to sources such as databases and web searches, and provide capabilities such as retrieval, conversation history, and agentic tool use — letting a model plan, act, and check its own work inside guardrails you build around it. The model performs the AI task; the system puts it to work.

Models

Generative models get most of the attention, but they are only one class among several. Embedding models, cross-encoders, vision-language models, and entity models are each designed for different kinds of work. Knowing the differences helps you choose the right model for the task — and often a smaller, faster, cheaper one.

So we teach with cards — one reference page for each model class and system pattern. They are yours to keep: enough to recognize what you're looking at, understand what it does, and choose the right tool long after class ends.

Start here

The systems that put AI to work

Each system below is built around one or more of the model classes further down this page, plus the code that connects them to data, to other tools, and to the people using them. Increasingly that includes agentic systems — ones that plan, call tools, and check their own work inside guardrails you build around them.

System Card 01

Agentic Workflow

A model that plans its own steps, calls tools, and iterates on the result instead of answering in one shot. The system defines what it's allowed to touch.

Built fromGenerative + tools

System Card 02

Chatbot

Interface, history, model, and orchestration. The model is stateless — the system is what remembers.

Built fromGenerative

System Card 03

Document Processing

Converts PDFs and scans into structured, searchable content — headings, tables, and reading order preserved.

Built fromVision-Language + parsers

System Card 04

Semantic Search

Finds passages by meaning rather than keywords, then reranks what it found.

Built fromEmbedding + Cross-Encoder

System Card 05

Retrieval-Augmented Generation

Retrieval plus generation: grounded answers with citations, drawn from sources you trust.

Built fromSearch + Generative

System Card 06

Model Evaluation

Candidates run side by side on the same tests. The goal is not a universal winner, but the right model for one use case.

Built fromMany models + a test set

System Card 07

Concordance Pipeline

Compares structured results from two independent sources to see whether they agree. Disagreement is a finding too.

Built fromExtraction + comparison

One layer down

The five model classes

Every system above is built from one or more of these five model classes — each with the architecture it is built on, what goes in, what comes out, and the models you would actually reach for — inside this course and beyond it.

Showing all 5 model classes

Model Card 01

Generative Model

The versatile engine behind chatbots, agents, and coding assistants. Writes natural language, code, and structured text from a prompt.

Architecture
Decoder-only transformer
Input
Text prompt
Output
Text
In the course · run locally GraniteGemmaPhi
Beyond the course Claude OpusClaude MythosGPT-5.6 SolGPT-5.6 TerraLlamaQwen

Model Card 02

Embedding Model

Turns text into vectors so meaning can be compared and searched. Text with similar meaning lands in a similar place.

Architecture
Encoder transformer
Input
Text
Output
Vector
In the course · run locally BGESapBERT
Beyond the course E5Nomic EmbedJinaOpenAI embeddings

Model Card 03

Cross-Encoder

Reranks results by reading the query and the passage together. It does not search — it judges what search already found.

Architecture
Encoder transformer, paired input
Input
Query + candidate passage
Output
Relevance score
In the course · run locally MS MARCO MiniLMMedCPT
Beyond the course bge-rerankerJina RerankerCohere Rerank

Model Card 04

Vision-Language Model

Reads images and document pages and answers in words. Gives generative AI the ability to see.

Architecture
Vision encoder + language decoder
Input
Image + prompt
Output
Text or structured data
In the course · run locally Gemma 4 E2BGemma 4 E4BGranite Docling
Beyond the course Claude OpusGPTQwen-VL

Model Card 05

NER Model

Labels the entities inside text — people, organizations, dates, diagnoses. Modern models take custom entity types described in plain language.

Architecture
Encoder transformer, span classification
Input
Text + entity labels
Output
Labeled entities
In the course · run locally GLiNERGLiNER2
Beyond the course spaCy NERSpanMarkerFlair

The through-line is judgment. Anyone can call an API. The skill worth having is knowing which model class the problem calls for, what it will cost you in speed and precision, and what has to be built around it before it is useful to anyone.

Every model on this page runs on your own hardware unless marked otherwise — no data leaves the building.

See the next cohort of Privacy-First AI Systems for Healthcare →