Vibeleaderboard
Index / app
Visit github.com
Category
AI Tools
Rank
No. 1423Tools index

Previous survey · No. 1377 ·

Type
APP
Use case
Models: Train & Run
Interfaces
Desktop
GitHub
77 stars
Latest release
v0.3.0
Date

About

A cross-platform desktop application and serving system for running Jev-like open decision models locally on your own computer. It is built with Tauri and PyTorch.

At a glance

Bud Decision Studio runs decision models locally and shows how they score the answer options you provide. It is worth trying if you want to compare structured choices without building your own model interface.

Good fit

Developers exploring decision models on supported desktop hardware.

Consider something else when

Use a validated process for consequential decisions where you need evidence that scores are calibrated on your own data.

Start here

Start with a fictional example with clear options, then reorder them and compare the scores. Repeat with labeled cases from your intended workflow before relying on a threshold.

Cost and requirements

The project documents builds for Apple silicon macOS, Windows and Linux. Models have their own licenses and hardware and storage needs.

Limits to know

The project documents local model execution and publisher-provided weights. Review the selected model’s terms and evaluate its scores before using them as probabilities in a workflow.

Read the source review

What it does

A decision model does not generate text. You hand it a situation, such as a support ticket or a log line, plus typed questions (pick one, yes or no, rate on a scale, pick any, put in order, estimate a number), and it returns a probability for every allowed answer in one pass. This app downloads eleven such open models, sets up the matching GPU or processor engine on first launch, and serves them through a desktop interface and a local HTTP server. It also keeps a local history of every decision, versioned templates, an evaluation page and a trainer for fine-tuning on your own labelled examples.

Why it's ranked here

It packages a whole workflow around a narrow model category: running, comparing, logging and retraining classifiers that return calibrated probabilities rather than prose. The pieces are specific. Installers ship for Apple Silicon Macs, Windows 10 and 11, and Linux on x64 and ARM64. The server answers on the same request format as TypeSafe's hosted decision API, so existing callers can point at it unchanged. The trainer publishes every run, including failed ones, with before and after accuracy on held-out examples. For a team that wants local, auditable routing or triage decisions, setup, serving, history and fine-tuning come in one install.

What's good

The trainer is cautious by design. It measures the released model first, mixes general examples into every batch, fits a calibration temperature, and keeps a fine-tune only if held-out accuracy rises and a separate set of general questions does not drop by 2 points or more. It then reloads the saved model and checks the answers match. The local server binds to this machine only, rejects requests whose Host header names another site, and refuses writes from other websites open in your browser. Each answer carries an act flag against a threshold, so code can act automatically or route to a person.

Tradeoffs

The quick install pipes a remote script straight into your shell, and the source setup does the same to fetch the uv package manager. Hardware support is uneven: AMD on Linux is experimental, the largest model needs an NVIDIA GPU, and training requires an RTX 30 series card or an M2 Mac or newer. Every published training result was produced by the project on one NVIDIA GB10 machine, so treat the gains as the authors' measurements. Every call, including the compatibility route, is stored in history unless you opt out. Streaming responses and multi-user workspaces are planned, not built.

How to use it well

Fits developers building routing, triage, moderation or labelling steps who want probabilities they can threshold instead of parsing free text from a chat model. Start in the Playground with a model's built-in example, then save the questions as a template and pin a version or alias from code. Use the act flag to send low-certainty answers to a human, and label those in History. Fine-tune only once you have a few hundred labelled examples with every answer represented; under 30 is refused. Set an API key before exposing the server beyond localhost. It does not generate text, so pair it with a language model for anything that writes.

Technical notes+

basal/server.py is a FastAPI app on 127.0.0.1 port 8420 that serves the UI, a TypeSafe-compatible wire route, OpenRouter and Vercel-shaped gateway routes, and the stateful studio API; models run in separate worker processes so the server never touches the GPU. Its middleware enforces a DNS-rebinding Host check, a required management header for state-changing calls, and Bearer auth with hmac.compare_digest when BASAL_API_KEY is set; without a key, remote callers reach only decision creation, preview and the model list. basal/guard.py blocks browser writes by Origin and Sec-Fetch-Site and requires JSON bodies from browsers on studio routes. basal/decisions.py runs resolve, check, run and record for every entry point and stores HMACs of sensitive variables under a per-install secret. docs/studio-api.md lists streaming, eval runs and comparisons as planned. install.sh fetches uv over curl and runs the engine installer.

Observed

Language
Python server (FastAPI, PyTorch); desktop shell built with Tauri
Platforms
macOS on Apple Silicon, Windows 10 and 11, Linux x64 and ARM64
Packaging
One-line install script, plus dmg, exe/msi, deb, rpm and AppImage downloads
Interfaces
Desktop app and a local HTTP API; TypeSafe-compatible decision route plus a stateful studio API
Accelerators
NVIDIA CUDA, Apple Metal, Intel XPU, AMD ROCm (experimental), CPU fallback
Default binding
Listens on localhost only; API key required for remote access to history and templates
Fine-tuning
LoRA trainer with held-out accuracy and forgetting checks before a model is kept

Read from README.md, requirements.txt, docs/trainer/README.md, docs/trainer/RESULTS.md, basal/server.py, basal/decisions.py, basal/guard.py, install.sh, docs/studio-api.md.

Tags

decision-modelsdesktop-applocal-aitauripytorchjev

Tech Stack

Python

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.