Brand safety · IAB 2.2 · Keywords · on ZeroGPU
The ad decision lands before the answer does.
Send a message on any of Koah’s four surfaces. Three specialized ZeroGPU models read it in parallel: Llama Guard 4 checks brand safety, the IAB classifier finds IAB 2.2 categories, and the signal model pulls keywords. Simple rules then pick an ad from a sample feed while GPT-OSS-120B writes the answer. Every model call is shown with its output, latency and cost.
Paste it into Claude Code, Cursor or ChatGPT to run these models on your own messages as a dry run, with the features your ranker would get. All it needs is a ZeroGPU API key.
Same decision, one general model
The comparison model gets one prompt asking for the same three things: brand safety, the top IAB 2.2 categories, and keywords copied from the message. Its answer goes through the same checks and the same feed rules.
Checking availability…
Cost per 1M ad decisions
From each call’s token usage at list prices, averaged over the messages you send on this page. The answer model isn’t included: it’s the app’s cost, not the ad network’s.
Send a message to see measured costs.
The ad feed
Fifteen fictional advertisers, each with declared IAB 2.2 categories and bid keywords. Three are in categories Koah’s content policy prohibits (alcohol, gambling, political), so they can win the match and still never serve. Review the feed to have the same models read every ad.
What each model does
Three ZeroGPU models make the ad decision, in parallel. Plain rules (no model) match the result to the feed. A fourth model writes the app’s answer.
Flags violence, weapons, hate, harassment, sexual content, self-harm and illegal activity. When it flags the message, there’s no ad.
$0.18 / $0.18 per 1M tokens · /v1/moderationsScores IAB Content Taxonomy 2.2 categories and audience segments. Every name is checked against IAB’s official 2.2 list. The main targeting signal.
$0.02 / $0.05 per 1M tokens · /v1/responsesPulls keywords from the message. The demo keeps the ones that appear in the message: 2 for a short message, up to 5 for a long one.
$0.04 / $0.10 per 1M tokens · /v1/responsesScores each ad on how its declared categories fit the message’s categories, plus keyword matches. Prohibited categories never serve. Below the threshold, no ad.
microseconds · lib/decide.jsWrites the app’s reply, streamed. It runs at the same time as the ad decision and isn’t part of it.
$0.15 / $0.60 per 1M tokens · /v1/chat/completionsThe whole integration
Three calls in parallel with your answer, then your own ranking. OpenAI-compatible endpoints, one API key.