# What is Jev? The AI that Can't Talk Back (and why that's a good thing)

## Executive summary

Jev is a new general-purpose AI classifier from TypeSafe that reads complex text and outputs only a single, predefined choice, eliminating the need for generative text. This capability allows developers to solve 'semi-deterministic' problems—those requiring judgment but yielding structured outcomes—at a fraction of the cost and speed of traditional Large Language Models (LLMs). The video outlines four architectural patterns for integrating Jev into existing software, positioning it as a crucial new primitive alongside deterministic code and LLMs.

## Key takeaways

- Jev's Core Functionality: Jev is a classifier that takes complex text input but is restricted to selecting from a predefined list of possible answers, making it fundamentally different from LLMs that generate sentences. This restriction is its primary advantage.
- Market Adoption and Value: Jev was the fastest-adopted model in Vercel's AI gateway history, suggesting that the ability to classify cheaply and quickly is a highly valuable, unmet need in the industry.
- The Power of Classification: Jev enables 'semi-deterministic' processes—where judgment is required (e.g., assessing customer dissatisfaction) but the output is a simple choice (e.g., 'High Risk,' 'Low Priority').
- Cost and Speed Advantage: Jev offers massive cost and speed gains compared to LLMs. For example, one developer reported a 34x cost reduction and six times speed increase when using Jev for tax document classification.
- Architectural Patterns: Jev can be used in four ways: 1) Classifying messy input for existing software; 2) Acting as a filter for large problem spaces (e.g., immunology research); 3) Choosing the next step in an AI workflow (outer loop orchestration); and 4) Interpreting user intent in real-time interfaces (e.g., spreadsheets).

## Technical details

- General Purpose Classifier: A classifier is simply a fancy word for multiple-choice selection. Jev is designed to be general-purpose, meaning it can be applied to diverse problems without requiring specialized retraining for each new category.
- Deterministic vs. Semi-Deterministic: Deterministic rules use ordinary code for predictable outcomes (e.g., 'if total > limit, flag'). Semi-deterministic problems require human-like judgment (e.g., 'Does the customer sound like they are about to leave?') but can be solved by classifying the output.
- AI Workflow Orchestration: Jev can be placed in the outer loop of an AI workflow, deciding whether the next step should be calling an LLM (for reasoning), calling a frontier model (for hard problems), or calling a human.
- Cost Efficiency: Jev's pricing is cited as 4.2 cents per million input tokens with zero charge for output. This makes it significantly cheaper than LLMs, enabling experimentation with large datasets.

## Practical implications

- Identify processes in your software that require interpreting complicated text but only need to output a simple decision (e.g., 'Yes/No,' 'Category A/B/C'). These are prime candidates for Jev implementation.
- Use Jev to build an outer loop for AI agents, allowing the classifier to decide the optimal next action (LLM call, human review, or tool execution) rather than relying solely on the LLM's internal logic.
- Implement Jev in data interfaces (like spreadsheets) to combine language judgment (e.g., 'Urgency') with traditional calculations (dates, amounts) to create richer, more intelligent data points.
- Test Jev against existing LLM calls for the same task (e.g., document classification) to quantify the cost and speed savings before migrating the function.

## Topics

Artificial Intelligence, Natural Language Processing (NLP), AI Architecture, Software Design Patterns, Machine Learning Classification, Jev-Shaped Problems Guide, Jev Classifier Setup Prompt

Source: https://www.youtube.com/watch?v=tYugqJ9YytQ