Jev Model

Product Introduction

1. Overview: The Jev Model is a structured decision-making AI model developed by TypeSafe AI. It operates as a deterministic reasoning engine that processes text or JSON inputs to produce typed, probabilistic outputs for application logic, distinct from generative text models.
2. Value: It enables developers to build reliable, auditable, and cost-effective decision automation by replacing brittle rule-based systems or complex prompt engineering with a single API call that returns structured data (Choice, Score, Noul) for direct integration into business logic.

Main Features

1. Structured Output Types: The model provides three core decision primitives. Choice selects from a predefined list of options (e.g., routing categories). Score assigns a discrete value from a defined scale (e.g., priority levels 1-5). Noul returns a boolean judgment against a specific criterion (e.g., "requires human review").
2. Decision Flow Engine: A visual canvas allows users to chain multiple decision questions into a single executable workflow (e.g., spam check → intent classification → routing). The API executes only the questions on the triggered path, optimizing cost and latency while encapsulating complex logic.
3. Batch Processing & Measurement: A dedicated console tool for uploading datasets (CSV/JSON) to run decision templates or custom questions at scale. This allows for pre-production validation, accuracy measurement, and threshold tuning on real data before deployment, reducing implementation risk.

Problems Solved

1. Challenge: Automating operational decisions like customer support triage, content moderation, or lead scoring often requires stitching together multiple AI services, managing complex prompt chains, and handling unstructured outputs, leading to high latency, cost, and maintenance overhead.
2. Audience: Product engineers, ML engineers, and operations teams at SaaS companies, marketplaces, social platforms, and customer support organizations needing to automate repetitive classification and routing tasks with high reliability.
3. Scenario: A support platform uses a Jev Decision Flow to ingest user messages. It first performs a Noul check for spam, then an intent Choice (billing, bug, feature request), and finally a priority Score. The entire logic runs in one API call, instantly routing the ticket to the correct queue with a defined urgency level.

Unique Advantages

1. Strengths & Limitations:
   • Pros: Predictable pricing per decision (not token), structured outputs eliminate parsing errors, visual flow builder for complex logic, and batch tools for empirical validation. The "unsure" branching feature automatically flags low-confidence decisions for human review.
   • Trade-offs/Cons: It is specialized for decision tasks, not for open-ended conversation or content generation. It requires clear definition of options and criteria upfront, which may need domain expertise.
2. Alternatives & Differentiation:
   • vs. GPT Function Calling/Structured Outputs: Jev is purpose-built for decision trees and batch analysis, offering a dedicated console, flow designer, and cost structure based on decisions rather than tokens, which can be more economical for high-volume classification.
   • vs. Traditional ML Classifiers: Jev requires no model training or ML infrastructure, allowing rapid prototyping and iteration via its UI and API. It leverages a large language model's reasoning but constrains it to a defined decision schema.

Frequently Asked Questions (FAQ)

1. What is the difference between Jev's Choice, Score, and Noul? Choice picks one label from a list (e.g., "spam", "legit"), Score assigns a numeric level (e.g., "priority 3"), and Noul gives a yes/no answer to a specific question (e.g., "Does this need escalation?").
2. How is pricing calculated for a Decision Flow? You are only billed for the Jev questions that are actually executed on the path through your flow. If a spam check (1 question) filters out a ticket, subsequent intent and routing questions on that branch are not called or charged.
3. Can I test Jev on my own data before integrating the API? Yes, the Batch Processing tool allows you to upload a CSV or JSON file with your texts, run any question or template flow against them, and inspect all structured answers and confidence scores to evaluate performance.