Bastion

Windows · 297 MB installer

SimpleML

Click-and-play machine learning, on your own machine.

Request SimpleML
The SimpleML Assistant page: a chat with a connected agent checking a tiny story model's project and runs, with a sidebar of projects.

Works with what you already run

Inside the app

The SimpleML Research page for a support ticket router project: a diagram of goal, orchestrator, two lanes, reward gate and best so far, with progress figures below.
A research run: a goal, an orchestrator that plans lanes, a reward gate and the best result so far.
The SimpleML Kit page for a tiny story writer project: a six-step setup from getting text to checking the model on text it has never seen.
The training kit: get text, set some aside, pick a design and train, then check it on text the model never saw.

One app instead of a stack of tools

Doing ML usually means stitching together notebooks, data-prep scripts, training recipes and eval code, each with its own setup. SimpleML is one local desktop app: connect an agent, drop in your data or start from scratch, and it prepares the data, trains, evaluates and shows you the results.

ONE APP · YOUR MACHINETraining kitspre-training and post-trainingResearch swarmsbuilt inYour agentCodex, Claude Code, local modelsEvaluationsealed test rows, scored hereYour datadrop it in, or start from scratch

How it works

Connect an agent, bring your data, and it prepares, trains, evaluates and shows the results.

  1. 01Connect

    Connect an agent

    Codex, Claude Code or a local model plans and runs the work.

  2. 02Data

    Drop in data, or start from scratch

    Bring your own files, or begin with nothing and let the agent set it up.

  3. 03Prepare

    It prepares the data

    The connected agent cleans and shapes the data for training.

  4. 04Train

    Pre-training and post-training kits

    Training runs from built-in kits, with research swarms for the open-ended search.

  5. 05Evaluate

    Honest evaluation on sealed test rows

    Test rows stay sealed during training, so the score you see is not inflated. The results are shown in the app.

Before and after

Measured on the previous app that SimpleML replaces: each change was tested on the tasks it was meant to fix and kept only if it beat the version before it.

Limits

The numbers on this page were measured on the previous app that SimpleML replaces, not on SimpleML itself. Ten runs a side is a small sample, and each fix tried alone did worse than the baseline; only the combination was kept.

Small ML tasks finished

Small ML tasks finished. Before 7 of 10, After 8 of 10. Higher is better.

10 runs per arm, small local model, Git Bash · 9 / 10 without Git · previous app. Source: lab/log.md, H-shell A5 and A6.

Failed shell commands per run

Failed shell commands per run. Before 2.0, After 0.0. Lower is better.

Same 10 runs · previous app. Source: lab/log.md, H-shell A5.

Time per run

Time per run. Before 119 s, After 85 s. Lower is better.

Same 10 runs · previous app. Source: lab/log.md, H-shell A5.

Tool router hits at 4 tools

Tool router hits at 4 tools. Before 101 of 140, After 132 of 140. Higher is better.

Of 140 steps, learned router vs BM25, p = 7.9e-9 · previous app. Source: lab/log.md, H1b.

Reports with an invented number

Reports with an invented number. Before 28 of 90, After 6 of 90. Lower is better.

Number check plus one repair call, p = 0.031 · previous app. Source: lab/log.md, E11.

Judge wrong keeps

Judge wrong keeps. Before 19 of 72, After 4 of 72. Lower is better.

With a plain text check, in-sample · previous app. Source: README, Checking the judge.

Prepare, train and evaluate, on your machine.

Request SimpleML