AN OPEN NOTEBOOK ON SYSTEM ONE AIRESEARCH NOTES · SEPT 2026

Software needs fast decisions.
Give it the right kind of intelligence.

“Software needs System 1 thinking, but we keep building System 2 thinking.”

A clear, practical guide to Jev AI. Understand how it works,
explore what people are building, and find your next useful idea.

01 / UNDERSTAND02 / EXPLORE03 / BUILDCURATED, CONNECTED, ALWAYS CURIOUS.
THE IDEA / BEFORE THE API

Two ways to think. Two kinds of work.

Explore all 44 slides ↓

In Thinking, Fast and Slow, Daniel Kahneman explains two modes of human thought: fast, automatic intuition and slower, deliberate reasoning. They work together. This distinction also offers a useful way to think about AI in software.

01System 1Fast & intuitive

React to a familiar situation.

On a familiar bike ride, you notice an obstacle and adjust almost immediately. Experience helps you respond without planning every movement.

IN SOFTWARE

Route this support ticket. Rank these passages. Choose the next valid browser action.

context → focused decision

Fast does not mean always correct. Intuition can be biased or miss important context.

02System 2Slow & deliberate

Work through an unfamiliar problem.

Planning the next five years of your career takes effort: assess your skills, compare opportunities, consider trade-offs, then build a plan.

IN SOFTWARE

Investigate a bug. Compare architectures. Plan a research task and explain the reasoning.

problem → analysis → plan

Deliberation takes more time and attention. It is valuable when the problem needs it.

LLM: understand & planJev: make focused judgmentsCode: validate & act

TypeSafe says Kahneman’s book inspired the name “System One Models.” The AI comparison is an analogy, not a claim that models think like brains. LLMs can also answer quickly; reasoning models are particularly useful for deliberate, multi-step work. Jev returns typed judgments that still need evaluation and checks.

THE 30-SECOND EXPLANATION

AI that gives your
code a decision.

Give Jev context and a question. Get back a choice, a score, or a yes/no probability your software can use directly.

Read the fundamentals ↗SOURCE: TYPESAFE DOCUMENTATION
CONTEXT → QUESTION → DECISIONFIG. 01
INCOMING SUPPORT TICKET

“I was charged twice.
Could you help with a refund?”

jev Which team?
Billing0.88
Technical0.08
Other0.04

Illustrative output · a typed answer, not generated prose.

THREE SIMPLE PRIMITIVES↗

Choice

Which option fits?

"billing"
≋

Score

How much, on your rubric?

2.4 / 3
◐

Noul

How likely is a yes?

0.92
↳

Let an LLM plan and write.
Let Jev make focused judgments.

OPEN SOURCE / BUILT WITH JEV

See what developers are building.

Explore on GitHub ↗

Real projects, the people behind them, and the code to explore. Search the directory, compare projects, or open a project for its README and developer profile.

148 projects from the community catalogue. Connecting to GitHub…

PUBLIC GITHUB API

Live GitHub results merge with the 19 Sep 2026 community catalogue. Every card labels when its numbers were checked. Illustrations are labeled separately from real demo previews.

Coverage, updates & image credits

This directory starts with 148 projects from Awesome Jev Use Cases. The public GitHub API searches for Jev in repository names and descriptions, 100 results at a time. Use “Fetch more from GitHub” to continue. GitHub limits each query to 1,000 results and may return incomplete matches; this is not a claim to list every project on the internet. Newly discovered matches are not independently audited.

Search responses are cached for 10 minutes; project and profile details for 30 minutes. No token is required for public data. API limits or errors leave the dated catalogue available. GitHub search documentation ↗

Real demo preview frames are from Awesome Jev Use Cases, with the original creator named. Other covers are original workflow illustrations, not screenshots of those products. GitHub supplies developer avatars. Source repository license ↗

PROJECT & DEVELOPER
01 / THE USE-CASE LIBRARY

What will you build?

Small, useful patterns. Real starting points.
Open a card to see how it works.

12 use cases

01
Agents

Route to the right model

Send a request to the handler or model that fits the job.

Choice
02
Trust & safety

Check an LLM’s work

Screen messages and tool calls before your application acts.

Noul + Score
03
Everyday tools

Turn reviews into ratings

Find which features a review mentions, then score each one.

Noul + Score
04
Search & data

Retrieve better context

Rank candidate passages before they reach the answering LLM.

Score
05
Trust & safety

Verify a citation

Check whether the cited passage actually supports a claim.

Choice
06
Agents

Give a browser fast decisions

Choose an action from the controls a browser can currently see.

Choice
07
Search & data

Extract without inventing

Choose the right value from candidates already found in a document.

Choice
08
Search & data

Screen a reading list

Check abstracts against your research inclusion criteria.

Noul
09
Agents

Keep useful agent context

Select relevant history when an agent’s context starts to grow.

Score
10
Search & data

Make text usable for ML

Turn language into signals for a downstream predictive model.

Score + Noul
11
Everyday tools

Triage a support inbox

Organize requests by team and urgency before a human opens them.

Choice + Score
12
Agents

Find an agent’s next skill

Select a relevant capability from a known skill catalogue.

Choice
12 CURATED STARTING POINTS

Patterns, community code and ideas are labeled inside each card.

02 / THE READING DESK

Go straight to the source.

The docs, code and research worth keeping.
A little context before you click.

START WITH THE BIG PICTURE

Understand the model.
Then question the claims.

Official documentation explains the interface. Community research tests the edges. Read both.

How we review sources ↗
Official[01]

Meet Jev, by TypeSafe↗

The model’s purpose, launch announcement and official links. Start with the maker’s own explanation.

TypeSafe AI
Reading note +

Product claims are vendor-reported. New signups were paused when reviewed on 25 September 2026.

Documentation[02]

The use-case map↗

A practical map of decisions: routing, retrieval, verification and more. Follow it into the worked cookbooks.

TypeSafe AI
Reading note +

Example workflows are starting points. Evaluate accuracy and thresholds on your own data.

Presentation[03]

Jev, explained in slides↗

A visual introduction to structured decisions, practical examples, architecture questions and limitations.

Shared research material
Reading note +

44-page slide deck, reviewed visually. Its “no independent benchmarks” slide is dated: JevBench now exists. Predictions are not established facts.

Code[04]

Smartphone review analyzer↗

A small Python example that turns reviews into topic ratings using Noul and Score questions together.

CampusX
Reading note +

Uses 50 synthetic reviews and a 0.5 mention threshold. Treat the threshold as a demo setting, not a universal rule.

Collection[05]

Awesome Jev use cases↗

A community collection of demos, repositories, reported results and useful limits, with original source links.

Boulanouar Walid
Reading note +

CC0 text and data. Social metrics are a 19 September 2026 snapshot; demo media belongs to its creators.

Collection[06]

What developers are shipping↗

A browsable directory of community projects, from SDKs and developer tools to experiments and agents.

shipwithjev
Reading note +

A discovery directory, not an independent audit of the listed projects.

Research[07]

Jev’s architecture, investigated↗

Black-box experiments explore shared state, question isolation and direct probability outputs.

Archer Hume
Reading note +

Independent analysis with explicit hypotheses. The proposed backbone and sparse architecture are not official specifications.

Benchmark[08]

JevBench: compare the evidence↗

Compare decision quality, calibration, latency and cost with a versioned evaluation and published methods.

Benchmark Heaven
Reading note +

Reviewed v1.4.2, scored 24 September 2026. Rankings depend on the task mix, weights, hardware and cost assumptions.

Research[09]

Laya: an open-weight alternative↗

Explore a different decision-model family, its checkpoints, multilingual routing and reported evaluation.

ConvAI Innovations
Reading note +

Author-reported results. Some comparisons mix fine-tuned and zero-shot models, or local and hosted latency. Self-hosting still has compute costs.

SOURCE MATERIAL / COMPLETE PRESENTATION

The Jev presentation, slide by slide.

Download PDF ↓

All 44 slides from the supplied presentation, in their original order. Use the arrows, jump to a page, or choose a thumbnail below.

▤ jev-presentation.pdf
Full size ↗
Slide 1 of 44 from the supplied Jev presentation
Browse all 44 slides

Source note: These are the original slides, not updated claims from this guide. In particular, the deck’s statement about the absence of independent benchmarks is no longer current; see JevBench ↗. Architecture hypotheses and reported performance should be read with their original context. Credit remains with the presentation’s author(s).

WATCH & LEARN

Study Jev AI with a video walkthrough.

Watch on YouTube ↗

Watch the explanation alongside the slides, and pause whenever you want to explore an idea.

03 / FROM READING TO BUILDING

One state. A few good questions.

Official quickstart ↗
1

Describe the state

Keep the relevant text and context. Leave unrelated information out.

2

Define the decision

Choose a primitive and make the options or rubric explicit.

3

Let code take it from there

Validate the answer and apply your thresholds. Review uncertain cases.

Use a server-side API key. Current access and model availability are listed in the TypeSafe console ↗.

example.js

Select one of your allowed options.

const response = await fetch(
  "https://api.typesafe.ai/v1/systemone",
  {
    method: "POST",
    headers: {
      Authorization: `Bearer ${process.env.TYPESAFE_API_KEY}`,
      "Content-Type": "application/json"
    },
    body: JSON.stringify({
      model: "jev-1.13.0",
      state: "I was charged twice. Please refund me.",
      questions: {
        decision: {
          "type": "choice",
          "instructions": "Which team should handle this ticket?",
          "criteria": {
            "billing": "Charges, invoices and refunds",
            "technical": "Bugs and outages",
            "other": "None of these or insufficient information"
          }
        }
      }
    })
  }
);
if (!response.ok) throw new Error(`HTTP ${response.status}`);
const { answers } = await response.json();
console.log(answers.decision);
04 / THE COMMUNITY SIGNAL

Around the Jev ecosystem.

Checking public sources…

What does this feed include?

Public GitHub projects, Hacker News stories and Bluesky posts are checked every five minutes while this page is visible. X is included when the server has an authorized API token. Unavailable sources are labeled; previously fetched results may be shown as stale for up to 24 hours.

Activity uses engagement and recency. It is a discovery signal, not a global trending chart. GitHub’s timestamps are repository push dates; stars are lifetime totals.

Keyword and source matching.

THE COMMUNITY ARCHIVE

Early experiments worth exploring.

From Awesome Jev · metrics captured 19 Sep 2026
A dated snapshot, separate from the fresh feed.

Browse all 74 archived demos +
Jev plays Doom@CompleteSkeptic ↗A canvas you control by pointing and speaking@jackcheng ↗Jev plays Subway Surfers@_MaxBlade ↗A real-time ad blocker@iam_zachi ↗500 emails for 3.5 cents@rileybrown ↗Jev plays Smash Bros. against itself@maubaron ↗Triage across 1,500 emails@ryanvogel ↗700 leads scored in 40 seconds@romanbuildsaas ↗Jev plays Super Mario Bros.@faadilhshaik ↗jev() for PostgreSQL@iam_zachi ↗Game levels generated in real time@HugoDuprez ↗Keystroke oracle@dabit3 ↗jevlike@vinnylarouge ↗1kpapers@nutlope ↗A model router on Jev@ephraimduncan ↗Computer use without screenshots@milindlabs ↗Every’s editorial vibe check@danshipper ↗400 companies matched to one candidate@sarvagya_kul ↗$10,000 in Jev’s hands@abolbuild ↗An always-on assistant with no wake word@_MaxBlade ↗Post scoring with SuperX@robj3d3 ↗Doomscroll Filter@robj3d3 ↗Predictive spreadsheets@dabit3 ↗The X algorithm, rebuilt with Jev@leojrr ↗SEO and GEO fixes, 90% cheaper@irabukht ↗Jev plays Slay the Spire 2@coolish ↗A chat bot with no LLM@CodingGarden ↗Simple Jev@picocreator ↗A Downloads folder that sorts itself@marcelpociot ↗Hide posts on X in plain language@marcelpociot ↗YouTube sponsor skipper@tdinh_me ↗End-to-end tests run by agents@o_kwasniewski ↗openjev-sglang@ekzhang1 ↗Live viral post analyzer@rileybrown ↗Intent-based search in Gmail@dabit3 ↗A second-hand shopping agent@AlanDaitch ↗Fraud detection with Jev and Kimi K3@nutlope ↗Ad creatives from filtered assets@higgsfield_ai ↗openjev on Qwen 4B@justALEXWORTEGA ↗Stagehand on a remote browser@kylejeong ↗3,282 posts, eight questions each@iannuttall ↗A visual reference finder@albicodes ↗TypeSafe Typewriter@stevekrouse ↗Lurk@mxfp4 ↗askjev.ai@waynesutton ↗jev-review@niazmorshed_ ↗900 images in 40 seconds@fayazara ↗A local Jev@wmoto_ai ↗Website to App@chddaniel ↗Jev Calc@thekitze ↗Jev Detector@jozef_gherman ↗An agent with a Jev model router@rileybrown ↗A filter for reply-guy comments@iannuttall ↗Headless Chromium agent@mormonnegro ↗Agentic browsing in Chrome@razaanstha ↗A prompt box that fills itself in@sawyerhood ↗A chief of staff for bots@milindlabs ↗Real-time Clippy@sotak ↗A Slack agent, twice as fast@johnyeo_ ↗Which outreach signals book demos@pierreeliottlal ↗DiffJury@raihankhan_rk ↗X timeline labeler@the_cyw ↗jev-job-hunter@hqmank ↗700 live ads in 40 seconds@Yarilo7brigada ↗Jev plays Tetris@AlanDaitch ↗One-click invoice finder@FarouqAldori ↗Jev as an agent safety monitor@isNickMa ↗AI slop detector@kraayenJon ↗
A NOTE TO BUILD ON

Start small. Measure what matters.

Pick one decision. Try it on real examples. Keep the evidence.
The most useful AI is the part that makes your software work better.

Find your first use case ↑

How it works

    Keep in mind

    Explore the source ↗