Link
https://github.com/gitcommitshow/jev-classifies-bots
Live site (optional)
No response
What does it do, and where does Jev fit?
What does it do?
A data pipeline transformation that classifies high-stakes events (signups, logins, checkouts, installs) as they flow through a customer data pipeline. For each event it answers two questions: what produced it (a person, an AI agent acting for one, a crawler, or other automation), and whether it's part of abuse. The verdict is written onto the event, so downstream tools can drop fake signups, keep agent-driven purchases in revenue models, and exclude bots from engagement metrics. It ships with shadow and enforce modes, a fail-open 2s timeout, and four test events with expected outputs.
Where does Jev fit?
Jev makes both decisions in one call per event: a Choice for traffic kind and a Noul for bad intent. They're asked separately on purpose. A bot farm running real browsers scores as human with high intent, and an AI agent checking out for a logged-in customer scores as ai_agent with low intent. User-agent rules get both cases wrong. Jev's latency is what makes this possible inline, inside a transformation's 4s execution limit, on live events rather than in a batch job afterwards.
Category
Applications
Link
https://github.com/gitcommitshow/jev-classifies-bots
Live site (optional)
No response
What does it do, and where does Jev fit?
What does it do?
A data pipeline transformation that classifies high-stakes events (signups, logins, checkouts, installs) as they flow through a customer data pipeline. For each event it answers two questions: what produced it (a person, an AI agent acting for one, a crawler, or other automation), and whether it's part of abuse. The verdict is written onto the event, so downstream tools can drop fake signups, keep agent-driven purchases in revenue models, and exclude bots from engagement metrics. It ships with shadow and enforce modes, a fail-open 2s timeout, and four test events with expected outputs.
Where does Jev fit?
Jev makes both decisions in one call per event: a Choice for traffic kind and a Noul for bad intent. They're asked separately on purpose. A bot farm running real browsers scores as human with high intent, and an AI agent checking out for a logged-in customer scores as ai_agent with low intent. User-agent rules get both cases wrong. Jev's latency is what makes this possible inline, inside a transformation's 4s execution limit, on live events rather than in a batch job afterwards.
Category
Applications