The course

Twelve notebooks. Each one runs from top to bottom in about a minute (notebook 02 takes about two), measures itself against labeled data, and prints what it cost. Every page here shows a real run with real outputs, so you can read before you run.

Pick your path

• 15 minutes: "What is this?"

  ---

  01 · Hello, Jev → 02 · Jev vs LLMs

  You'll know what Jev does, how fast and cheap it is, and where it breaks.

• 1 hour: "Put it in an agent"

  ---

  01 → 03 · The agent loop → 04 · Email triage → 07 · Guardrails

  You'll plug Jev into a real agent and build two tools you can use today.

• A weekend: "End to end"

  ---

  All 12, in order, finishing with the capstone.

  You'll have built all 13 use cases and a complete ops copilot.

All lessons

#	Lesson	What you learn	What you build
01	Hello, Jev	the three question types, many questions in one call, JSON state, criteria	your first calls, Jev vs an LLM head to head
02	Jev vs LLMs	accuracy, latency, cost, calibration; Jev's known limits	a fair benchmark and a chart
03	The agent loop + Jev	the five decision points: router, Jev as a tool, guard, done gate, judge	an agent that refuses to shut down a core network link
04	Email triage job	batch classification, confidence routing, LLM only where needed	a scheduled inbox-triage job (works with real IMAP)
05	SMS scam shield	combining code signals with Jev; three-way verdicts	a scam-text checker for your family
06	Code vulnerability hunter	chunk with ast, map with Jev, reduce with the LLM	a security scanner that found 9/9 planted bugs
07	Guardrails	action guards, human approval, prompt-injection shields, failing closed	a shell agent that can't be tricked into leaking data
08	Model and tool router	tier routing with live prices, tool shortlists	a router that cuts cost by about half
09	On-call log triage	templating in code, Jev per template, SEV from your runbook	an incident summary from raw logs
10	RAG relevance and citations	relevance filtering, citation checking	a docs Q&A bot that says "I don't know"
11	Jev as judge	Jev vs LLM judges, CI eval gates, measuring gates	an eval suite that runs on every PR
12	Capstone: ops copilot	everything together	a copilot that routes 25 mixed items in ~9 s

How every notebook is built

The same structure repeats on purpose, so each new notebook only teaches what's new:

1. Config cell: one OpenRouter key for both brains. It's identical in every notebook.
2. Helpers cell: ask_jev() for the fast brain, chat() for the slow brain, and SPEND, which tracks the exact cost of every call.
3. The agent loop (where one is needed), identical everywhere, from build-your-first-ai-agent, plus two hooks.
4. The lesson: data, questions, measurement against labels, and the recap.

Download and run

Each notebook page has a download button at the top. To run notebooks locally, follow Start in 3 steps.