Skip to content
afsah
← All work

Decision model + desktop assistant

Jevlet

What should happen next?

A small model that chooses among available actions, joined to a Windows assistant that shows its interpretation and asks when confidence is insufficient.

State
Research implementation
Working with
Python / PyTorch / BERT encoders / Windows UI Automation

01 The constraint

A desktop action needs more than plausible language: it needs a valid target, a bounded action and a decision about whether to proceed. Jevlet reconstructs publicly described System-One ideas as an independent research project. It does not claim to reproduce TypeSafe’s proprietary Jev implementation.

02 The decisions

Choose from what exists.

The model returns probabilities over runtime-defined options rather than generating an action as text. Installed apps, open windows, visible controls and spans of the command become candidates. This makes the available choices explicit, though it also means a missing candidate cannot be selected.

Share state. Keep questions separate.

One packed sequence encodes the desktop state and multiple question branches. The attention mask lets each branch see the state and itself, while hiding sibling branches. The option-isolated variant also keeps options separate until the decision token compares them.

Keep a research track beside the daily driver.

The scratch backbone remains useful for studying topology without pretrained representations. The desktop track uses a pretrained encoder under the same topology. The assistant exposes confidence, alternatives and a demonstration path so corrections can become local training data.

  1. 01Desktop state
  2. 02Candidate choices
  3. 03Typed decisions
  4. 04Confirm or execute
A bounded decision path. Confidence and risk influence whether the assistant acts.

03 What exists

The repository contains the decision API, scratch and pretrained backbones, calibration and benchmark tooling, and a Windows assistant. Its interaction model includes alternatives, confirmation for risky or uncertain requests, and local demonstrations for personalization.

04 Still open

A choice model cannot act outside the options it receives. Public Jev details do not establish its unpublished internals, and Jevlet’s design choices are its own. The documented laptop benchmarks are narrow evaluations, not evidence of reliable automation across arbitrary applications. New training recipes and version plans remain distinct from completed evaluation results.

Follow the evidence

Based on project code and documentation reviewed in October 2026. Evaluation results are reported by the projects; they were not independently rerun for this portfolio.

Next projectOMNI