Jev and the Judgment Layer

Two talks in one evening.

Talk 1. Jev: a model that answers with a probability by Tom McLaughlin. Jev, by Typesafe, is a narrow model. It does not write text. It answers typed questions with a probability: is this email relevant, which team does this page concern, does this source contradict this claim. It cannot reason or look things up, so every fact it needs must be in the question. In return, it is fast and cheap enough to run on every item.

The talk shows how to write questions that a narrow model can answer, and how to set the thresholds for "act", "skip", and "ask a person". It shows where Jev replaces a call to a large model in a working pipeline, and where a narrow model is the wrong tool.

Talk 2. The Judgment Layer, by Brad Risse. Companies record their money, their code, and their deals. They do not record their judgment: the decisions people make every day in meetings, email, and chat. This talk is about that missing record.

An agent proposes an action. A person approves it, rejects it, or edits it. The judgment layer stores each decision in a ledger. The difference between the proposal and the final decision becomes evaluation data, and the agent improves from it. As the agent earns trust, it gets more autonomy, one workflow at a time.

The talk shows what a judgment ledger stores, how decisions become evaluation data, and how to decide when an agent can act without approval. The example is a working knowledge base that reads email, Slack, calendars, and meeting transcripts.

After the talks. At 8:00 PM, we go to Cunningham's Pub & Grill (Old Mill), 10904 W Dodge Rd, near the DevObsessed office, for an informal social. Everyone is welcome, including people who cannot make the talks.

Free to attend.