Context
Claire Vo hosts How I AI, a show about practical AI workflows. This episode is a solo walkthrough of Jev, a new model from TypeSafe AI that returns structured decisions instead of generated text. It matters to PMs because most product work involves triage, classification and routing decisions, and Claire argues a model built only for that job is fast and cheap enough to run on data sets that were previously too tedious or too costly to analyze at all.
The Big Idea
Jev is not a chatbot or a coding model. It is a decision model that returns a choice, a score, or a yes/no probability instead of text, and because it is extremely cheap and fast, it makes classifying huge messy data sets practical for the first time.
Claire says she spent under $10 on Jev tokens over a week and used it to build analysis she believes is worth hundreds of thousands of dollars. The value comes from pairing a cheap, fast classifier with a smarter model for the reasoning step, not from Jev doing everything alone.
Key Insights
Jev returns values, not text
Every normal LLM takes text in and returns generated text out. Jev takes text in (including text descriptions of images, though not images directly) and returns one of three type safe values: a choice picked from a predefined list, a score such as a severity ranking, or a nool, a probability that the answer to a yes/no question is true. Claire's example: asked "is Claire a podcaster," Jev would return something like a 99% nool rather than a sentence.
Jev is priced for high volume
Jev costs 4 cents per million input tokens and charges nothing for output tokens, since it barely outputs anything. Claire ran it on 112 PRs from her marketing site for 1.1 cents, and on roughly 1,700 PRs from her ChatPRD app for 9 cents, finding 17,000 matched pairs in about two minutes. That cost profile is what makes it worth pointing at data sets nobody would normally pay to analyze.
Pairwise comparison is Jev's best trick
Rather than asking Jev to name a category directly, Claire feeds it pairs of items (two PRs, two comments) and asks only "are these related, yes or no." Repeating that across a data set builds clusters without ever asking Jev to invent a label. She then hands those clusters to a cheap generation model (Gemini Flash Lite) to name the themes. On her ChatPRD repo this surfaced that almost 30% of engineering effort went to platform security and infrastructure, a number she says leadership and boards ask for constantly.
Local coding sessions hide data
Claude Code and Codex sessions are stored locally, so Claire ran the same classification approach on her own session history. In January she was almost entirely doing engineering tasks; by September, less than 40% of her sessions were engineering, with more time going to agent work, video publishing, and a new business line. She frames this as meta-analysis anyone can run on their own machine, at close to zero cost.
Pair Jev with a smarter model
- What: Claire's working pattern is to use Jev to tag, cluster, and filter a large corpus, then apply a stronger reasoning model only to the clusters that matter.
- Why it matters: Jev alone can classify but not explain; a big model alone is too slow and expensive to run over everything, but is exactly right once the data is already narrowed down.
- Example: on her Gmail inbox, Jev scored subject lines and snippets for whether an email could be deleted, then handed the remaining email to another model to actually process.
A cheap filter unlocked a product
Claire had spent thousands of dollars trying to build a "product insights graph" for ChatPRD using frontier models alone, and called it "just hard" because of how much nuance was in the raw data. Splitting the pipeline by job (Jev for classification, clustering and mapping; a stronger model like Astra for analysis and strategy; a generation model for writing) let her process about 1,100 raw signals (PRs, support tickets, meeting notes, linear tickets) into over 200,000 classifications and pairings for about $4 in Jev costs, plus more on the reasoning step. She says this made the feature actually work where brute-forcing it with big models alone had not.
Speed makes real time apps possible
Because Jev is fast as well as cheap, it can sit inside a live loop instead of only running in batch. Claire built a live search over 4,500 YouTube comments and a voice-to-color app that maps spoken sentiment to a hex color and a matching quote in near real time, using OpenAI's realtime voice API to capture speech, Jev to score sentiment against a list of colors, and a quotes API to fetch a matching quote. She's explicit that a fast demo like this still needs real architecture underneath, such as batching results and pushing high scores to the top, not just calling Jev on everything at once.
Mental Models & Frameworks
The choice, score, nool trio
Jev's entire output surface is three primitives. A choice selects one item from a list you supply (what should I wear: a dress, jeans, workout clothes, ski gear, or nothing). A score rates something on a scale, such as triaging a bug as cosmetic, broken, or blocking. A nool is a probability that a yes/no answer is true. Claire's framing: this covers roughly 90% of what software engineering actually does day to day, which is why it generalizes to so many use cases.
Route each job to its model
Claire's product insights pipeline assigns each step to the model built for it: classification, clustering, and mapping go to Jev; analysis, strategy, and insight extraction go to a bigger reasoning model (Astra); generation goes to a model built for writing (Soul or Luna). The lesson is to decompose a hard problem into steps first, then match each step to the cheapest model that can actually do it, instead of routing everything through one expensive model.
Practical Application
Run a PR classification pass
Pull your repository's pull requests through Jev using pairwise "are these related" comparisons to build clusters, then label the clusters with a cheap generation model. Claire's version cost 9 cents on about 1,700 PRs and produced a percentage breakdown of engineering effort by initiative, the exact number leadership usually has to ask for manually.
Analyze your own local coding sessions
Point the same pairwise classification approach at your local Claude Code or Codex session history to see how your actual time allocation has shifted over the year, engineering versus agent work versus other tasks. This costs close to nothing and needs no new data collection, since the sessions are already stored on your machine.
Build a cheap triage layer first
For any large, messy inbound stream (support tickets, emails, comments), use Jev to score or classify first (severity, sentiment, deletable or not), then only send the flagged or high-value subset to a more expensive reasoning model. This mirrors Claire's Gmail workflow, where Jev decided what could be deleted and a second model only had to handle what remained.
Batch results before scaling a classifier
Before wiring a Jev-based classifier into a live feature, batch and rank the results rather than scoring an entire data set on every request. Claire's live comment search over 4,500 comments only felt instant because results were pre-batched and high scores were pushed to the top, not because Jev itself removed the need for that architecture.
Bottom Line
Jev is a narrow, extremely cheap decision model that returns a choice, a score, or a probability instead of text, and its real power shows up when it is used to classify and cluster large data sets before handing the narrowed result to a smarter, more expensive model. Claire's examples, from PR analysis to a stalled product insights feature, all follow the same shape: let Jev do the sorting, let a bigger model do the thinking.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Jev (TypeSafe AI) | Decision model that returns a choice, score, or probability instead of text | The subject of the episode; 4 cents per million input tokens, no output charge |
| Gemini Flash Lite | Low cost generation model | Used to label the clusters Jev built from pairwise comparisons |
| Astra | Reasoning and analysis model | Handles the strategy and insight extraction step in Claire's product insights pipeline |
| OpenAI Realtime Voice API | Real time speech capture | Captures Claire's voice for the emotion to color demo |
| API Ninjas Quotes API | Quote lookup by filter | Returns a matching quote for the voice to color demo |
| YouTube Data API v3 | Pulls video comments | Used to pull 4,500 How I AI comments for the audience signal dashboard |
Notable Quotes
"Jev alone is okay. Jev with an LLM buddy is super powerful." (Claire Vo)
"I think I cracked this baby with Jev." (Claire Vo)
