Jev: 8 real use cases this fast, cheap model

In this episode of "How I AI," host Claire Val welcomes back AI engineer John Lindquist to discuss Jev, a newly released decision model that has captured their attention over the past week. Unlike traditional large language models that generate unstructured text, Jev takes unstructured input (natural language) and returns structured, type-safe outputs with confidence scores and limited predefined options.

John demonstrates eight practical use cases for Jev. First, he shows a voice-controlled to-do app where users can dictate tasks like "make an appointment with the dentist, go buy oat milk, complete request check, low priority" and Jev classifies the information in real-time while the user is still speaking, determining when enough data exists to execute a function call. The system performs multiple passes—verifying transcription accuracy, confirming task appropriateness, and matching to the correct operation—all instantaneously.

Second, John presents document and contact reconciliation, where Jev analyzes large datasets to identify duplicate or similar records (e.g., "Cedar Grove Office Products" matching "Cedar Grove Office"). This pairwise comparison works on massive scales with confidence scoring, allowing users to set thresholds for when merges should occur, and can process 60,000 or 600,000 objects in milliseconds.

Third, he demonstrates a command router or omnibar that intelligently directs user queries to the appropriate application or tool within a larger system. Rather than users remembering specific tool names, natural language queries are routed to the correct subsystem, which then executes the requested action.

Fourth, Jev is shown playing chess competitively, evaluating all possible moves, ranking the three best options, then evaluating subsequent moves, all within one second—10 times faster and four times cheaper than using traditional LLMs for the same task. This demonstrates how speed enables novel applications like real-time games.

Fifth, John shows a Wikipedia pathfinding demo where Jev navigates from any topic (e.g., LeBron James) to the Philosophy article by analyzing links, embodying goal-directed routing.

Sixth, a warehouse agent coordination system shows Jev managing three simultaneous agents with different tasks, ensuring they don't collide or interfere with each other, checking at each step whether tasks can be executed safely in parallel.

Seventh, the presentation trainer use case monitors a speaker's real-time speech against a prepared list of discussion points, checking off items as they're mentioned and alerting the speaker if time is running out and key points remain uncovered.

Eighth, John mentions YouTube comment clustering where Jev groups comments, ranks them, and enables real-time search across thousands of comments—a task prohibitively expensive with traditional LLMs.

Throughout the discussion, Claire and John emphasize Jev's pricing (approximately 4 cents per million input tokens, with some providers offering free access) and execution speed. They contrast Jev with LLMs: Jev excels when decisions involve a limited, defined set of options (routing, classification, ranking), while LLMs are better for open-ended, creative, or generative tasks where the output space is unlimited.

A key insight emerges: Jev enables "effective inefficiency," where computing all possible options, ranking them, and checking for conflicts is computationally intensive but now economically feasible, unlocking use cases that were previously deemed too expensive or slow to be worthwhile. John notes this is why developers feel more in control with Jev—they can see and reason about the decision paths, similar to traditional if-else programming logic.

Both speakers discuss limitations: Jev sometimes requires multiple passes or classification refinement, and users may initially misunderstand when to use single-pass vs. multi-pass approaches. However, since Jev is so cheap and fast, adding verification layers is practical.

The conversation concludes with John discussing broader AI trends: the challenge of local storage as developers work with multimedia and 3D files across Mac minis, and the paradigm shift toward agent-centric workflows (like Grok bots) where agents become the primary organizational unit rather than files and folders. Claire mentions running 40 different Grok bots across her business. John is working on mega.dev, an educational program teaching how to build, evaluate, and implement AI agents effectively.