Using Decision Models for Voice, Routing and Data

Claire Vo interviews developer educator John Lindquist about demonstrations using Jev, a decision-oriented model. The examples are exploratory applications shown in September 2026, not independent benchmarks; the episode has a sponsor.

The hosts describe Jev as returning constrained choices and scores rather than open-ended prose. Lindquist suggests using those outputs at decision points within a larger application, with ordinary code and other models handling the rest.

A voice to-do demo classifies spoken commands to add, complete or remove tasks. Multiple passes help match imperfect transcription to an existing task and select an action.

The pair discuss chaining classification decisions: each step narrows the options for the next. This lets an application map natural language to a known function instead of asking a model to produce an arbitrary program.

Other examples include mapping grocery requests to functions and proposing merges of duplicate records. They mention confidence thresholds and a second model for review; a model’s score is not automatically a calibrated probability, and consequential merges still need validation.

An application router can choose a destination and action from a limited set. Vo adds that a responsive search experience still depends on architecture, such as grouping and ranking candidates, rather than assuming inference has zero latency.

They use chess and other constrained tasks to discuss when a decision model may be appropriate. They contrast this with open-ended brainstorming, where a generative model may be more useful.

Lindquist shows pathfinding and multi-agent coordination examples, including avoiding collisions in a grid. These are demonstrations of a particular setup, not evidence of general reliability in physical or high-stakes environments.

A presentation coach listens for whether a speaker covers predefined points. The hosts discuss testing different formulations and composing multiple classifications when one decision is insufficient.

They close by reflecting on how people design and supervise multiple agents. Their comments about organizational roles are observations and predictions, not established employment outcomes.