AI/ML and Data Science

Day 8–10 · The Welzin edge

Welzin builds with AI. That doesn't mean every engineer trains models - it means every engineer should be able to integrate a model into a product, evaluate whether it's actually working, and reason about its failure modes.

The instinct we're training: treat an LLM like an untrusted, occasionally-brilliant intern. Verify, scope, monitor.

This chapter is split into four sub-chapters. Work through them in order. Each ends in hands-on checkpoints - your buddy will review them before you mark the parent chapter complete.

How this chapter is structured

Data Science 101 - the classical ML toolbox. When stats and a linear model beat an LLM (more often than people admit).
GenAI 101 - the model hierarchy, prompting as software engineering, embeddings, structured outputs.
RAG Pipelines - retrieval done seriously: chunking, hybrid search, re-ranking, citations, evals.
Agentic Workflows - tool use, the agent loop, multi-step planning, safety patterns, when to reach for the Claude Agent SDK.

Each sub-chapter is a self-contained primer with its own hands-on checkpoints and PDF/reading resources. The grid below will take you there.

Sub-chapter map

#	Sub-chapter	What you'll be able to do
1	Data Science 101	Pick the right model class for a problem; build a baseline before you LLM it
2	GenAI 101	Write a prompt you can defend; use embeddings; ship structured outputs
3	RAG Pipelines	Build a RAG that actually retrieves the right thing, with citations
4	Agentic Workflows	Wire tools to a model and not get burned by it

A note on order

If you're new to the field entirely, do them 1 → 4. If you're already comfortable with classical ML, you can skim Sub-chapter 1 and dive into 2.

You do not get to skip Sub-chapter 4 because it's "boring infra." The agentic safety patterns there are the difference between shipping a delight and shipping a footgun.

What we don't cover here

This chapter is the operating manual, not the theory book. We point at the seminal papers and the Anthropic / Eugene Yan / Hamel Husain writeups in each sub-chapter's further-reading slot - you should read those, but you don't need to before you can be productive on a Welzin codebase.

The next chapter (Chapter VII - Claude & Skills) is the platform-specific complement: how to operate Claude as a tool at Welzin level of fluency.

Welzin opinion: Anyone can wire up an LLM call. The engineering is in evals, retrieval quality, cost discipline, and failure handling. That's where we win or lose customers.