T3 · Required case
Livestream labeling with TypeSafe Jev
Paste comments → Jev Choice suggests a label. Humans decide what to pin, reply, or ignore. High confidence → auto-suggest; low → human queue. No money or refund decisions.
Policy: AI suggests only. Never auto-refund, never capture payment. No accuracy promises — calibrate thresholds on your own labeled chat.
Labels
| Key | Meaning | Example |
|---|---|---|
| buy_intent | Ready to buy / stock / shipping before or during purchase | “In stock? Can you ship to the US?” |
| price_question | Asking about price, discount, or deal terms | “How much for two?” |
| after_sales | Post-purchase support | “Package arrived damaged.” |
| spam | Bots, ads, scams, harassment, noise | “DM me for free followers!!!!” |
| other | Residual / unclear | “lol fire stream” |
Paste demo path
- Operator pastes one or more comments into your app (batch as separate evaluate calls or structured state).
- App calls typesafe-ai/jev via AI Gateway with the Choice question below.
- UI shows suggested label + probabilities; low confidence stays in a human queue.
- Operator applies the tag in their tools — Jev never moves money.
Question definition
{
label: {
type: 'choice',
instructions: 'Label this livestream comment for the host team.',
criteria: {
buy_intent: 'Purchase intent, stock availability, or shipping destination questions tied to buying',
price_question: 'Asking about price, discounts, bundles, or deal terms',
after_sales: 'Post-purchase support, returns, defects, delivery problems after buying',
spam: 'Bots, ads, scams, harassment, or irrelevant promotional noise',
other: 'Does not fit the above',
},
},
}Illustrative paste → suggestion
Static demo only — this site does not call the model.
Pasted comments
1. In stock? Can you ship to the US?
2. How much for two hoodies?
3. DM for crypto signals!!!
4. Item arrived torn, need exchange
Suggested labels
1. buy_intent (high → auto-suggest)
2. price_question (high → auto-suggest)
3. spam (high → auto-suggest)
4. after_sales (high → auto-suggest)
Low confidence → human queue
Action: present suggestions; human confirms.
Confidence policy (example only)
- High confidence — High confidence — auto-suggest the label in UI (still no financial side effects).
- Low confidence — Low confidence — leave unlabeled / send to human queue.
Recalibrate on your own labeled chat exports. These bands are illustrative — not accuracy promises and not defaults to ship blindly.