Jev Use Cases: What 562 Demos Built, and What Shipped

TypeSafe AI launched Jev on September 15, 2026, and builders flooded X with demos. Catalogs of them are not scarce: the awesome-jev list by Li-Evan on GitHub tracks more than 3,400 projects. One smaller gallery did something the big lists do not. The Jev demos collection on GitHub Pages (maintained by the GitHub account webdevcody) gathered 562 demos from launch week, September 15 to 21, and asked Jev itself to label each one by domain, by what was actually built, and by how useful it is. That turns a highlight reel into a dataset.

Read as a dataset, it says something the listicles miss. Classification and agent demos turn into products. Games, robots and trading bots collect views and stay demos. Of the 562, 9 are something a person can go and use today.

What is Jev, in one paragraph?

Jev is a decision model, not a chatbot. Code sends it some context and a list of questions with fixed answer shapes (pick one option, give a score, or say yes or no), and Jev returns the answers with a probability attached to each. It cannot write sentences, which is exactly why it is fast and cheap: the 2026 TypeSafe launch post, Introducing System One Models and Jev, pitches it as a function call for judgment. The price and the gap between the advertised and measured savings are covered in this site’s Jev cost breakdown. This page is about what people actually built with it.

How the 562 demos break down

The gallery’s labels come from Jev (model jev-1.13.0), not from this site. Each demo gets one domain, one form (prototype, open source, benchmark, product or commentary), and a utility tier running from toy through niche and useful to valuable. That is a model grading its own fan mail, so treat the tiers as a rough sort, not a verdict. The counts below were pulled from the gallery’s embedded data on September 24, 2026.

Domain	Demos	Rated useful or valuable	Live products
Agents (browser, computer use, tool choice)	132	85 (64%)	1
Classification (routing, moderation, labeling)	118	91 (77%)	6
Games	92	2 (2%)	0
Research and benchmarks	62	21 (34%)	0
Simulations and robotics	34	3 (9%)	0
Finance and trading	30	2 (7%)	1
Voice and media	28	10 (36%)	0
Consumer apps	26	8 (31%)	1
Business workflows	20	12 (60%)	0
Commentary about Jev	20	2 (10%)	0

Most posts never got past the video. 378 are one-off prototypes, 77 released code, 74 are benchmarks with numbers, and 24 are opinion posts with nothing built.

How many Jev demos reached a real product

Stage	Tasks remaining	Percent of candidate pool kept
Posts in the gallery (September 15 to 21, 2026)	562	100%
Built or measured something (excludes 24 opinion posts)	538	96%
Released code or a product (77 open source + 9 products)	86	15%
Live product (something you can use today)	9	2%

The domain split tells a clearer story than the funnel. When the bars are normalized so every domain has the same length, the useful work bunches up at the top.

Jev demo utility by domain

Category	Valuable	Useful	Niche	Toy	Total
Classification	32	59	25	2	118
Agents	11	74	42	5	132
Business workflows	6	6	7	1	20
Voice and media	2	8	8	10	28
Research	14	7	36	5	62
Consumer apps	1	7	11	7	26
Robotics and sims	1	2	21	10	34
Finance	1	1	23	5	30
Games	0	2	49	41	92

The numbers quoted below are what each builder reported in their own post. None were rerun by this site, with one exception: the Search Console triage under use case 4, which I ran on my own site. Most compare against a single frontier model on one task. Read them as order-of-magnitude evidence.

Use case 1: sorting and routing things at volume

This is where Jev’s shape fits best, and where most of the products come from. The pattern: one fixed question, asked of every item in a big pile, with the answer written to a database or used to pick a path.

• Routing customer messages in production. Elliot Trabac, an engineering manager at Gorgias, tested Jev in production behind a model router that picks a model for each customer message from a handful of typed questions. The router sits in Cortex, the company’s internal AI agent. His framing is the useful part: much of what gets shipped as “agentic” is a finite set of decisions with a model narrating between them.
• Classification inside SQL. MotherDuck shipped prompt_jev(), a SQL function for text classification. On 100,000 rows the company reports 40 seconds and $0.50, against 32 minutes and $37 for a frontier LLM. A separate open-source DuckDB extension, duckdb-jev, from Prasanth J reports 2,049 rows in 0.887 seconds.
• Scam and spam screening. Cellou Diallo built a WhatsApp scam radar that read 10,208 messages in 32 seconds for $0.065. Ryan Vogel ran Jev over 1,500 of his own emails, and three of the nine live products are Chrome extensions that score or hide X posts as you scroll.
• Document triage. Trinay Hari reports Jev classified a 26-sheet construction plan set in 2.9 seconds for $0.0052, matching the GPT-4.1 pipeline his company Proq already ran.
• Picking which AI model runs. Higgsfield uses Jev to read each prompt and choose which image or video model on its API should handle it. Jev does not generate anything; it only routes.
• News and brand matching. Elvis Sun’s brand news matcher read 384 morning headlines and told 15 brands which to act on in 24.9 seconds for $0.19. He reports Claude Opus 5, run on the same feed at the same time, finished 4 of 384.

Use case 2: faster browser and computer-use agents

Agents are the largest group (132 demos) and the one with the most open-source code (41 repos). The shared trick: turn the screen into a list of possible actions, have Jev pick one, and hand only the typing to a small language model.

• Browser agents. Gregor Zunic released a small open-source browser agent built on Browser Use that found flights in 7 seconds for $0.0039, with Jev picking the next action and target each step. Kyle Jeong paired Jev with Stagehand for remote browser control and reports $0.001 for the task shown. Cline built a jev-browser plugin into its desktop app, the one agent entry that is a live product.
• Desktop apps with no API. Stanislav Sorokin’s jev-clerk booked 30 real supplier invoices in a desktop accounting app using the real mouse and keyboard. Jev picks one action per step from a closed list, and every 10 invoices Claude Fable 5.1 reads the decision log and rewrites Jev’s playbook.
• Phones. Somesh Sampat’s open-source Android agent opened YouTube and subscribed to a channel in 32 seconds and 17 steps, reading the accessibility tree rather than screenshots.
• Cheaper tool calls. Rasswanth put Jev in front of Lyzr’s tool-calling loop to decide whether a turn needs tools at all and to shrink the tool list sent to the main model. Tamara Tran used it for instant context compaction: score every tool call in an agent session and drop the irrelevant ones, instead of running a summarization prompt.

This is the same move described in the site’s harness engineering guide: keep the loop in code, and spend a large model only on the steps that need words.

Use case 3: business workflows and back-office triage

Only 20 demos, but 60 percent rated useful or valuable, second only to classification. These are narrow jobs a person does by hand today.

• Sales pipelines. Ronald ran Jev over notes on 8,690 cold CRM deals in 7 minutes 41 seconds for $0.43. It sorted out 4,422 that had already said no, 300 with a passed callback time, and 633 worth calling, ranked.
• Hiring. Mukesh Bishnoi scored 1,000 resumes against one job description in 1 minute 28 seconds, rating skill depth, domain fit and seniority before filtering on hard requirements.
• Contracts and due diligence. Chris Poensgen deployed a contract repository demo, hosted at jev.eigenweltlabs.com, that asks one question across a set of documents and returns labeled answers with probabilities.
• Compliance alerts. AJ Asver, co-founder of Grep, describes a harness that moves repeated steps from LLM calls into code, with Jev making the decisions. His estimate for 100,000 compliance alerts: more than $290,000 on Opus 5, under $26,000 with the harness. Runtime Wire’s coverage of the AgentRun launch puts the per-alert figure at $2.89 on Opus 5 against $0.25, from an internal Grep test on 100 alerts.
• Branching forms. Tamir built JevForm, a form that chooses its next question from the previous answers instead of hand-written if/then rules.

Use case 4: checking other AI systems

The quietest group is also one of the most practical. A cheap judge makes it affordable to check everything instead of a sample.

• Scoring every production trace. LangChain shipped Jev-as-a-judge in LangSmith, pitched as scoring every production trace rather than a sample, including safety and security checks.
• Agreement with frontier judges. Benedikt Stroebl reports that Jev as the judge in Harbor Rewardkit reached perfect agreement with Fable and GPT on the Harvey LAB benchmark while running 50x faster. For how LLM judges are scored in general, see the site’s guide to LLM evaluation metrics.
• Code review. Paolo Rosson’s PR reviewer returns 14 typed checks (hardcoded secret, SQL injection, deleted tests and more) in one call, at a reported $0.00007 per PR, or 7 cents per 1,000 PRs against about $14.50 on Opus 5.
• Content and SEO audits. Borja rebuilt the internal link map for a 586-page site in 45.1 seconds for $0.21: 584 links placed, and 139 pages left unlinked because nothing honestly fit. Steven Michael Thomas ran a 1,000-page content audit through both Opus and Jev and reports Jev 75x cheaper and 25x faster.
• Search Console triage (tested on my own site). This is the one use case on this page I have run myself. Jev judged 1,161 query and page pairs from one of my sites with five yes or no questions, for 281 calls and under five cents, and Claude Code fixed what it flagged. Over the next weeks the site’s average Google position went from 12.7 to 6.2, queries ranking in the top 10 went from 49 to 1,518, and clicks went from 1.6 a day to 248. The full workflow, code and before-and-after data are in how I used Jev for SEO with the Search Console API.

Use case 5: real-time loops (games, robots, voice)

These are the demos that got the views, and the ones Jev rated lowest. They prove speed, not business value: a decision has to land in a fraction of a second or the game is lost.

• Games. The launch thread’s Doom demo ran at about 10 calls per second, which TypeSafe founder Diogo Almeida puts at roughly $7 an hour. Ronak Malde’s Jev and Astra setup beat the Minecraft Ender Dragon in 8 minutes 43 seconds for under $1, of which $0.01 was Jev. AI/ML API played blitz chess with Jev against Fable 5.1 and GPT-6 Astra, one API call per move.
• Robots and simulations. Simeon Li had Jev triage 300 incidents in a simulated 10,000-robot warehouse at a median 0.53 seconds per decision. Zyvex flew a simulated drone with Jev choosing over, around or brake at 2.5 decisions per second, with a local safety veto underneath. Leo’s city traffic-light demo reports average wait rising more than 600 percent when Jev is switched off.
• Voice. Ashish Bhatia’s healthcare voice agent hears “I’m having chest pain”, stops the booking flow, and routes to clinical staff in 500 milliseconds for $0.0004. It is the one voice demo that doubles as a safety gate. Andy Gao’s voice-controlled Mac opens an app before the sentence finishes, and Tharusha Kavishan’s open-source Svara lets blind and low-vision users browse the web with Sinhala voice commands.
• Consumer apps. Nailthy Tang’s real-time virtual try-on for Drape picks an outfit from a closet in about 620 milliseconds for $0.0011 per decision. Chris Tate’s json-render experiment has Jev compose a product’s own UI components on the fly.

What the demos warn you about

Finance is the most instructive group, because it is where confident wrong answers cost money immediately. Jev rated 28 of the 30 trading demos niche or toy.

The clearest lesson came from Adrian Cortex, who ran a Jev trader on MON/USDC that had to buy or sell every block, with no option to abstain. Jev was 85 to 88 percent confident on nearly every call. After 1,838 decisions in under 10 minutes it was down 5.2 percent. Sam Silva found the same gap on ordinary tasks: in a 1,000-item check across five public datasets, answers Jev gave at 70 to 80 percent confidence were right only 54 percent of the time, so he acts only on answers above 90 percent.

Both point to the same rule. Jev’s value depends on what happens to the answers it is unsure about. Throw them away (or force a guess) and the confidence score is decoration. Escalate them and the cost advantage shrinks by exactly the share escalated. The cost breakdown works through how much a 30 percent abstain rate costs: the saving is capped at 1 divided by the abstain rate, however cheap Jev gets.

On the other side, Arne’s Bitcoin replay shows why speed can matter on its own. On the same 258 trades, filling at Jev’s 250 milliseconds earned 631 basis points for the day against 207 at GPT-6 Astra’s 2.6 seconds. That is a latency result, not proof Jev picks good trades.

Which Jev use cases are worth building?

Across the 562 demos, the ones that became products share four traits:

1. The answers are known in advance. A queue name, a yes or no, a score from 1 to 5. If the output needs a sentence, a language model still does that part.
2. The same question repeats thousands of times. Price only matters at volume. One judgment a day is cheaper to make by hand.
3. A wrong answer is recoverable. Routing, labeling and flagging for review all survive mistakes. Forced trades and physical control do not, which is why those stay demos.
4. There is somewhere to send the doubtful ones. A person, a bigger model or a default action. Without it, the confidence score does nothing.

If a workflow passes all four, the demos above give a realistic sense of speed and cost. If it fails the third or fourth, it will make a good video and a bad product.

Frequently asked questions

What can you build with Jev?

Anything that asks the same fixed question many times: message and ticket routers, spam and scam filters, document classifiers, lead and resume scoring, AI output judges, and the action picker inside browser or computer-use agents. It cannot write text, so anything that needs a written reply still needs a language model.

What are the best Jev use cases so far?

Classification and routing. Of 562 community demos posted September 15 to 21, 2026, Jev rated 77 percent of classification demos useful or valuable, and 6 of the 9 live products are classifiers or routers.

Which companies use Jev in production?

Public examples from launch week include Gorgias (a message router in its internal AI platform), LangChain (Jev-as-a-judge in LangSmith), MotherDuck (the prompt_jev SQL function), Cline (a jev-browser plugin) and Higgsfield (model routing on its API).

Is Jev good for games or trading bots?

It is fast enough to play games and trade in real time, but those demos rarely become products. Jev rated 90 of 92 game demos niche or toy, and one trading demo that could not abstain lost 5.2 percent over 1,838 trades despite 85 to 88 percent confidence.

How were the 562 Jev demos labeled?

The community gallery at webdevcody.github.io/jev-demos collected 562 launch-week demos (larger lists track over 3,400 projects) and had Jev itself (model jev-1.13.0) assign each demo a domain, a form such as prototype or product, and a utility tier from toy to valuable. They are a model rating demos of itself, so treat them as a rough sort.

Sources

webdevcody (2026). Jev demos: 562 Jev demos, labelled by Jev for creativity, wow factor and utility. Community gallery, GitHub Pages. Data read September 24, 2026. https://webdevcody.github.io/jev-demos/

TypeSafe AI (2026). Introducing System One Models and Jev. Company blog. https://typesafe.ai/blog/introducing-system-one-models-and-jev

Li-Evan (2026). awesome-jev: The most complete gallery of what people build with Jev. GitHub repository. https://github.com/Li-Evan/awesome-jev

Runtime Wire (2026). Grep launches AgentRun, says code-first harness cuts agent costs 91%. News coverage (secondary). https://runtimewire.com/article/grep-agentrun-compliance-agent-costs-jev

Gorgias (2026). Cortex. Company product page. https://cortex.gorgias.io/

Individual demos are cited inline to the builder’s original post on X. Each was checked against the post text on September 24, 2026, and all figures are the builders’ own reports, not independent measurements.