Claire Vo’s Jev Experiments: Structured Decisions for Product Work

Claire Vo introduces a week of experiments with TypeSafe AI’s Jev, a model designed to return structured decisions rather than long generated text. The episode is sponsored by OpenArt, and the product and pricing details reflect September 2026.

Vo says Jev has prompted ideas for personal productivity, engineering and product work. Her estimate of the value of that work is a personal assessment, not measured revenue or a controlled productivity result.

She contrasts general language models that generate text with a decision-oriented interface that returns constrained values. Both can process text, but the output format and intended workload differ.

Vo explains that pricing can make high-volume classification attractive. Her quoted rates and subsidized access are historical; current pricing, limits and data-handling terms should be checked before using a service.

The outputs she demonstrates include choosing among options, scoring and estimating probabilities. A structured value is easier for software to consume, but its format does not guarantee the decision is correct or calibrated.

Her first example classifies pull requests by the type of work they represent. She uses the results to see how engineering effort is distributed across areas such as maintenance, content and tooling.

She reports low costs for runs on her repositories. Those are her own measurements for particular inputs and the pricing available at the time, not a universal cost estimate or accuracy benchmark.

Another example analyzes locally stored Claude Code and Codex sessions to categorize work. Before submitting such records to a service, the user must consider secrets, personal information, permissions and the service’s data policies.

Vo describes scoring email snippets for triage, then using another model to work with the categories. She does not display the private messages and recommends the combination of a decision model with a language model for richer analysis.

For ChatPRD, her own product, she builds an insights graph from signals such as pull requests, support tickets and conversations. The model helps classify and group signals; another model then helps interpret the groups.

She describes processing around 1,100 signals and many pairwise classifications. Grouping and generated interpretations should be checked against original records, especially when the output informs product decisions.

Vo demonstrates a dashboard built from thousands of YouTube comments. It groups feedback and displays categories such as sentiment and episode ideas, making the source material easier to explore rather than replacing it.

A second demo maps live speech to colors and quotes. She describes how a real-time speech service supplies text, Jev scores or selects among colors, and another API supplies quotes.

She notes latency in one part of the demo and encourages looking at the architecture behind apparently instantaneous applications. Speed depends on the complete pipeline, not solely on the decision model.

Vo closes by summarizing three patterns: analyze data, pair structured decisions with a generative model and build interactive experiences. The cases demonstrate her own experiments, not a general comparison against every model.