Question 1

Can existing Jev client code talk to it?

Accepted Answer

Yes. The README says the studio speaks TypeSafe's Jev API and gateway formats, so the official Python and JavaScript SDKs work by pointing base_url at the local server, by default http://127.0.0.1:8420.

Question 2

Does it train on my own decisions?

Accepted Answer

Yes. The Train page takes a spreadsheet of past decisions, recommends a model and runs LoRA on local hardware, keeping the new model only if it improves on held-out examples without regressing general decisions.

Question 3

How do I install the app?

Accepted Answer

One line on macOS Apple Silicon and Linux with curl piping get.sh into sh, or PowerShell with irm into iex on Windows. Or download the build for your machine: an aarch64 DMG, an x64 setup executable or MSI, deb packages for Ubuntu or Debian on amd64 or arm64, rpm for Fedora or openSUSE, or an AppImage for anything else. The binaries are published from the Bud-Decision-Engine releases page.

Question 4

What happens on the first run?

Accepted Answer

The app checks your graphics, processor, memory and free disk, then asks whether models should run on the GPU or the processor and states what each option installs and how large it is. It installs its own private Python with the PyTorch build matching your hardware and every model library, inside its own folder, and downloads nothing until you pick models, with two suited to your machine pre-ticked and downloaded one at a time in the background.

Question 5

Which hardware path will my models take?

Accepted Answer

NVIDIA GeForce, RTX, GB10 and data-centre cards run through CUDA 12.6, 12.8 or 13.0 chosen from your driver; Apple Silicon goes through Metal; Intel Core Ultra or Arc graphics use XPU; AMD GPUs on Linux use ROCm, which is experimental; anything else falls back to the processor, where small models answer in about a second. You can switch between GPU and processor later on the System page.

Question 6

What does a decision response actually contain?

Accepted Answer

Each answer says whether it is certain enough to act, measured against the template's act_threshold, so your code can either proceed automatically or ask a person. The decisions endpoint returns act, needs_review listing the questions that fell short, the answers themselves, and the template id and version. The Playground's JSON and Code tabs show the exact request that produced it.

Question 7

How do templates and versioning work?

Accepted Answer

A template is a reusable decision: its questions, the variables that fill the situation, a default model and settings. Every save creates a new version, so code can pin support-triage@3, follow a moving tag like support-triage@production, or take the latest. History then keeps every decision from the Playground or any program on this computer, and you can label the right answers and rerun on another model.

Bud Decision Studio

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community