Use your own data

Every use case is built from two things you can swap: the data and the questions. The loop, the helpers and the patterns stay the same.

Swap the data

Use case	Point it at	Where
Email triage	your mailbox (read-only IMAP)	.env → IMAP_*, or load_imap() in notebook 04
Scam shield	your helpdesk's "is this real?" reports	SMS list in notebook 05
Vulnerability hunter	any Python repo	ROOT in notebook 06 (the sandbox root)
Log triage	your service logs	LOG_DIR and LINE_RE in notebook 09
RAG filter	your docs, as Markdown with ## sections	load_passages("your/docs") in notebook 10
Judge / evals	your Q&A pairs with reference answers	data/eval_set.jsonl format

Rewrite the questions

This is where the real work is, and it's worth doing carefully:

1. Start from the decision you'd make by hand. "Would I reply to this?" is a better question than "Is this important?".
2. Describe every option. "billing": "payments, invoices, refunds, charges" is better than "billing": None.
3. Add a none or other option so Jev isn't forced to pick something that doesn't fit.
4. Use criteria on Noul questions to say exactly what counts as yes.
5. Label 30-50 examples of your own data and measure accuracy before you trust it. Every notebook shows how.
6. Route low confidence to a human. It's the cheapest way to make a classifier safe.

Keep it safe

• Keep tools read-only until you've added a guard and a human-approval step (notebook 07).
• Keep hard rules in code (phishing is never answered; SEV1 always pages).
• Fail closed when a check can't run.