Lesson 04 · 6 min

Use cases

Anywhere your software makes a repetitive judgment — classify, route, score, or gatekeep — is a candidate for Jev. Here are the patterns showing up in production.

Ticket triage & routing

Score every ticket for urgency, owning department and customer frustration in parallel — in milliseconds, with full coverage instead of random sampling.

Example · Login issue → account department, confidence 1.0

Agent tool-call guardrails

Screen risky tool calls (like bash) before they run and block the dangerous ones. This 'Auto Mode' pattern uses a System One model to insure the System Two model.

Example · Dangerous shell command → blocked before execution

Model routing

Send each request to the cheapest model that can handle it — fast model for lookups, powerful model for hard reasoning — cutting overall inference cost.

Example · Simple lookup → fast model · Architecture task → powerful model

Refund & compliance review

Put the ticket, the order and the refund policy into one state. Jev judges whether a refund was requested and whether policy supports it.

Example · Refund requested 0.99 · policy supports it 0.98

Account risk scoring

Score account activity on a risk scale and attach a calibrated recommendation to escalate — auditable by design.

Example · Risk score 1.84 · escalate probability 0.81

High-frequency agents

Because judgments cost fractions of a cent, whole new agent styles become viable: browser agents, live trading bots and email triage at scale.

Example · ~$7/hour to run a real-time game-playing agent

When not to reach for Jev

The task needs open-ended reasoning, writing, or generating new information — that's an LLM job.
The judgment criteria themselves require understanding complex, novel context you can't define up front.
Your state is larger than the 32K-token context window (chunk or summarize first).
You haven't validated calibration on your own data yet — regression-test probabilities against historical cases before trusting thresholds in production.

Where you can run it today

Cloudflare Workers AI

typesafe/jev

OpenRouter

typesafe/jev-1.13

LangChain

langchain-typesafe

TypeSafe console

console.typesafe.ai (early access)

The starter play

Pick one low-risk, high-volume decision — email or ticket triage is ideal. Mistakes have a human safety net, and the payoff is immediate. Once you trust the calibration, add agent guardrails on any tool that runs bash, sends email, or touches a database.

PreviousHow it worksNextTry the sandbox