Skip to Content
View all events

Building with Jev and System One Models

Published by O'Reilly Media, Inc.

Intermediate content levelIntermediate

Understand the model, explore use cases, and evaluate fast decisions in Python

Course Outcomes:

  • Explain System One thinking and Jev’s decision interface, including how it differs from LLM generation, structured outputs, and conventional classification
  • Use Choice, Score, and Noul in Python to classify requests, score urgency or frustration, and combine atomic judgments with business rules
  • Evaluate a Jev workflow against an LLM baseline and use validated thresholds to route uncertain cases for review
  • Identify suitable use cases, limitations, and relevant SDKs, agent integrations, and community resources

Many AI applications need a decision rather than a generated paragraph: which team should handle a request, which tool an agent should use, or whether a case needs review. This two-hour course introduces TypeSafe AI’s Jev and the System One approach to fast, structured decisions. Learn what the model does, what TypeSafe has disclosed about its design and training, and how it differs from conventional classifiers, small language models, reasoning models, and LLM structured outputs.

Using Python and a companion GitHub repository, you’ll build a support-triage workflow with Jev’s Choice, Score, and Noul primitives. You will add review thresholds and compare decisions, latency, token usage, and cost with an LLM implementation. You’ll also examine practical applications and the emerging ecosystem of SDKs, agent integrations, adapters, and community patterns. Throughout, you’ll distinguish schema validity from decision accuracy and vendor claims from measured results.

This live event is for you because...

  • You’re a developer or AI engineer exploring Jev and alternatives to using generative LLMs for every task.
  • You work with agents, support automation or decision workflows and want to understand where specialized decision models fit.
  • You want practical experience comparing model behavior and trade-offs before adopting a new model.

Prerequisites

Basic Python programming and familiarity with JSON, API calls, and LLM applications

No prior Jev experience or model-training background required

Recommended preparation:

  • Install Python 3.10 or later and an editor or notebook environment
  • Clone the companion GitHub repository and install its pinned dependencies before class (link to come)
  • Configure a TypeSafe API key or OpenRouter key and, optionally, an API key for the comparison LLM (API usage may incur charges)
  • Run the setup check using the supplied sample data

Course Follow-Up:

Schedule

The time frames are only estimates and may vary according to how the class is progressing.

Understanding Jev and System One models (20 minutes)

  • Presentation: System 1 and System 2 as an analogy; Jev’s typed decision interface; parallel outputs; TypeSafe’s disclosed training approach and what remains undisclosed
  • Discussion: Which tasks need text generation, a decision model, or ordinary code?
  • Q&A

Comparing Jev with existing approaches (20 minutes)

  • Presentation: Conventional classifiers, small LLMs, reasoning models, JSON mode, and constrained structured outputs; separating interface differences from accuracy and calibration claims
  • Demonstration: Applying the same ticket-triage questions to Jev and an LLM; inspecting answer spaces, probability distributions, and output behavior
  • Q&A

Building typed decisions in Python (20 minutes)

  • Demonstration: Preparing state and defining Choice, Score, and Noul questions with the Python SDK
  • Hands-on exercises: Classify a ticket, score frustration, and estimate whether it expresses urgency; inspect distributions and distinguish probability from the Choice and Score confidence statistic
  • Q&A
  • Break

Composing and evaluating a workflow (30 minutes)

  • Presentation: Keeping deterministic policy in code; selecting review thresholds using labeled cases and the consequences of errors
  • Hands-on exercises: Combine Jev judgments into a support-triage workflow; route clear cases, flag uncertain ones, and run the same dataset through the LLM baseline; use recorded responses if API access is unavailable
  • Group discussion: Compare decision accuracy, latency, token usage, cost, and review rate; inspect mistakes, including confident errors; separate benchmark observations from general claims
  • Q&A

Use cases and the emerging ecosystem (10 minutes)

  • Presentation and demonstration: Intent and tool selection, RAG relevance, output checks, and operational triage; reviewing official SDKs, the System One LLM adapter, coding-agent skills, and independent Jev Patterns examples; where Jev complements a writer or reasoning model
  • Group discussion: Choose one application and define the evidence needed to adopt Jev; discuss community experiments, current limitations, and which ecosystem developments to watch

Wrap-up and Q&A (10 minutes)

Your Instructor

Apurva Misra

Apurva Misra is a machine learning engineer, AI consultant, speaker, and founder of Sentick. She works with teams building practical AI systems, with a focus on moving from prototypes to reliable production workflows. Her work includes agentic systems, evaluation, context design, and the operational patterns required to make AI systems usable in real organizations.

Skill covered

QA / Testing