fastcampus-jev
Korean tutorial: 13 notebooks and a sample app attaching Jev to DeepAgents for tool choice, guardrails and risky-action gates.
Overview
This Korean tutorial, seminar material from Fastcampus by TeddyNote, teaches a split where Jev makes judgments and an LLM writes sentences, using a customer-support agent for a sample shopping mall built on DeepAgents (LangGraph). Thirteen notebooks and a React plus LangGraph sample app walk through the pattern end to end, with probability bars shown for every Jev judgment.
Topics include tool selection (Jev narrows tool candidates with a choice question and the LLM fills only the arguments), guardrails (injection, profanity, defamation and personal data judged in one call, including hidden instructions in tool results), a risky-action gate that asks whether the user directly requested an order cancellation or refund and hands ambiguous cases to a human, four open-source patterns (memory compaction, tool gate and output judgment, search and model routing, browser action selection), and document RAG where Jev picks folders and files, grep finds passages, and Jev decides whether to answer and whether the answer is grounded. Jev is called through OpenRouter.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which tool should handle this customer request? | — |
| Yes / no | Does this message contain injection, abuse, defamation or personal data? | yes · no |
| Yes / no | Did the user directly request this cancellation or refund? | yes · no |
How it works
- Customer message enters the DeepAgents agent
- Jev runs guardrail checks and narrows tools
- LLM writes tool arguments and replies
- Risky operations pass a Jev gate or go to a human
Strengths
- Comprehensive, hands-on curriculum
- Visual probability bars per decision
- Covers routing, guardrails, RAG and gating
Caveats
- Korean-language material
- Requires OpenRouter credit for Jev
Getting started
git clone https://github.com/teddylee777/fastcampus-jev.git, uv sync, set OPENROUTER_API_KEY, uv run langgraph dev --no-browser, then npm run dev in frontend.At a glance
- Type
- Article
- Status
- prototype
- Language
- Python
- Runtime
- Jupyter notebooks + LangGraph + React app
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Thirteen notebooks and a sample app on attaching Jev to DeepAgents for tool choice, guardrails and gates on risky operations. Korean.
MrJev/awesome-jev
Frequently asked questions
What framework does it use?
DeepAgents on LangGraph, with a React front end.
How is Jev accessed?
Through OpenRouter with prepaid credit.
What language is it in?
Korean.
Related projects
OpenRouter cookbook recipe gating agent tool calls with deterministic checks plus Jev Noul probabilities and fixed thresholds.
ArticleproductionTypeSafe's official cookbook for a two-stage skill-suggestion workflow that selects a skill and can reject the shortlist.
ArticleproductionMIT plugin and local MCP server that hands an agent's closed questions to Jev or local Laya, with policy gates and receipts.
Agent skillproductionPythonCognitive control layer for OpenCode that uses Jev to compute GO, DELIBERATE, VERIFY or BLOCK for agent actions, observing by default.
ProjectbetaTypeScriptPlugin suite adding Jev judgments to DeepSeek Harness: tool pruning, dead-loop blocking, a safety guard and skill routing.
PluginbetaTypeScriptjevkit: zero-dependency CLI and eleven MCP tools giving agents typed Jev decisions for routing, triage, guarding, grep, rank and compaction.
MCP serverbetaJavaScript