Building and Fine-Tuning Decision Models for AI Agents

AI applications need to route requests, prioritise messages, and decide when to ask for human review. Decision models give you a way to turn context and explicit criteria into structured answers your application can use.

In this workshop, you’ll work through the process from integration to adaptation: build with Jev, explore local alternatives, and fine-tune Qwen3 on a small labeled dataset. Throughout the session, you’ll compare outputs, investigate mistakes, and evaluate whether each approach fits the task.

What you’ll leave with

• Eight practical Python examples: Model routing, tool gating, results reranking, bulk labeling, real time control, LLM evaluation, Guardrails, and inbox triage, with decision criteria you can modify.
• Experience with local decision models: run Kev-4B and Laya localy and inspect their outputs on shared examples.
• A fine-tuning workflow: prepare labeled data, separate training and evaluation examples, and adapt Qwen3 for decision tasks.
• An evaluation approach you can reuse: compare baselines and checkpoints, inspect disagreements, and identify what needs further testing.

The three-hour agenda

00:00–00:30 — Jev and the decision layer

State, questions, criteria and typed answers. Eight use cases fit into an agent workflow, and how application code validates results and handles review and execution.

00:30–01:00 — Build routing and inbox triage

Work through the two Python labs. Change inputs and criteria, inspect outputs, and compare the decisions with simple rules and a structured-output LLM baseline.

01:00–01:15 — Short break

01:15–01:45 — Run Kev 4B and Laya

Run open local decision models on shared examples. Inspect model-specific outputs and establish a baseline before training.

01:45–02:30 — Fine-tune Qwen 3 for to become decision model

Prepare a small labeled dataset, separate training and evaluation examples, and run a compact fine-tuning exercise. Prepared checkpoints support the comparisons, so the session does not depend on a long training run finishing.

02:30–3:00 — Compare checkpoints and discuss your use case

Evaluate base and fine-tuned variants on the same examples. Inspect errors and disagreements, discuss what to test before deployment, and bring your questions.

Who this workshop is for

Software engineers and aspiring AI engineers who want to build with decision models and explore local adaptation.

You should be comfortable with basic Python, JSON and running a script or notebook. Previous Jev experience is helpful but not required. No previous decision-model fine-tuning experience is needed.

Reserve your seat

• Early bird: $49 until 1st of November. Standard price: $69
• Date: Sunday, 22th November, 19:00 - 22:00 EEST
• Location: Live on Zoom

Bring your questions and a decision task from your own application.