At a glance
WHAT IT’S REALLY ABOUT
How Jev enables real-time, low-cost decision-driven product features
- Claire Vo and John Lindquist explain why Jev (a fast, cheap decision model) changes what product teams can build by turning natural language into structured choices, scores, and function calls instead of text.
- They demo real-time voice-driven interfaces (a to-do manager and a presentation coach) that rely on sequential Jev calls to validate streaming input, match entities, infer intent, and execute actions without pauses.
- They highlight high-ROI data operations—deduplication, record merging, clustering, and large-corpus discovery—made feasible by Jev’s low latency and very low token cost, especially when paired with confidence thresholds.
- They introduce Jev as an application/tool router that can sit above many smaller tools (or models) to route users to the right workflow, while noting that search-like experiences still require architecture choices such as clustering, caching, and batching.
- They contrast Jev with traditional LLMs using a chess benchmark and UI/DOM examples to show that decision models excel when the possible actions are constrained, while LLMs remain better for open-ended interpretation and creative analysis.
IDEAS WORTH REMEMBERING
5 ideasJev is best understood as a “decision/function layer,” not a chatbot.
Unlike chat-oriented LLMs (text-in/text-out), Jev is framed as unstructured input → structured decisions (function choice, labels, scores). This makes it especially useful for connecting natural language to APIs, tools, and deterministic product actions.
Speed + near-zero cost unlocks “previously not worth it” workflows at scale.
Multiple demos emphasize sub-second responses and extremely low costs (e.g., cents for many runs, ~4¢ per million input tokens discussed). That changes what’s economically feasible—like scanning huge JSON corpuses or running lots of comparisons that previously felt wasteful.
Real-time UX comes from chaining small, fast Jev calls—not one giant prompt.
The to-do voice demo chains several Jev passes: validate dictation quality, match text to an existing task, then match intent to an operation (remove/complete/priority). It can also detect when enough information has arrived mid-stream to safely execute a function call.
Use Jev to propose merges with confidence scores, then validate/explain with a second model.
For messy databases (duplicate contacts/companies/password entries), Jev can do pairwise matching and propose merges quickly, with confidence thresholds (e.g., only merge if ≥99%). A follow-up model can audit samples or generate human-readable rationales (“why these two are the same”).
Jev works well as an ultra-fast router for apps, tools, and even models.
They describe building Jev into a multi-level router: first pick the correct app/tool (e.g., “to-do app”), then pick the specific action inside that tool, enabling “go to X and do Y” flows. This extends to model routing (choosing which downstream model/tool to invoke).
WORDS WORTH SAVING
5 quotesFast and free are both amazing. Um, if you've ever been tired, uh, sending a basic request to an LLM and waiting around for a bit just to have it, like, call a function, this is now essentially instant.
— John Lindquist
This model, which is basically free and incredibly fast, allows you to do discovery over data in a way that feels like it's opening up my opportunities and allowing me to, like, look at things that I thought weren't high ROI before.
— Claire Vo
I think of LLMs being unstructured to unstructured—where you put text in, you get text out. And this one is similar. You put in unstructured sentences and data, but then you get structured data out.
— John Lindquist
Whenever you have constrained inputs like, uh, that are interacting with apps, it's, uh, Jev is a, a good thing to, to reach for.
— John Lindquist
Jev unlocks efficient inefficiency, which is like it is kind of like inefficient to map out all possible routes, rank them, double-check that they're not gonna collide.
— Claire Vo
High quality AI-generated summary created from speaker-labeled transcript.
