Skip to content

Data Science Skills by Probabl

Tell us your problem. Let's experiment together.

Bring your agent and your model. Our skills and libraries organize, build, and evaluate the project, with the expertise from more than a decade of building scikit-learn.

Get started Book a demo

How to get started

Your agent and your model.
Our skills.

  1. Install our CLI

    Use the package manager you already have.

    ❯ pip install skore-cli
    ❯ conda install -c conda-forge skore-cli
    ❯ uv tool install skore-cli
    ❯ pixi global install skore-cli

  2. Install our skills

    Add our skills to your coding agent.

    ❯ skore skills install

  3. Start your agent

    Open the coding agent you already use.

    ❯ pi
    ❯ opencode
    ❯ claude
    ❯ codex

How it works

From your data science problem.
To deeper experiments.

  • 01 Do it right from the start

    Setup up your project

    We do the heavy lifting for you: scaffolding a proven data science workspace, managing a reproducible Python environment with your favorite package manager, recording every dependency, and keeping the project under version control.

    • A structured workspace ready for new experiments and production

    • Dependencies and experiment history preserved with Git

    Terminal showing project setup choices for a Python environment, workspace layout, and git

  • 02 Get the right insights

    Explore your data

    We dig into every data table: inspecting, analyzing, and exploring it to surface insights and catch issues early, researching ideas from the literature, and turning every finding into code for a detailed report you can return to.

    • Insights and data issues caught before any modeling

    • A detailed report with code you can come back to

    Browser window showing an exploratory data analysis report with a California housing value map

  • 03 Get evidence you can trust

    Build and evaluate models

    We take it from data source to prediction: connecting the tables with skrub, modeling with a scikit-learn-compatible library, and evaluating in skore with a structured report of the right metrics, issue checks, and insights on the data, pipeline, and results.

    • Any data source, one scikit-learn-compatible pipeline

    • A skore report of metrics, checks, and insights

    Browser window showing a linear regression report with a SkrubLearner pipeline diagram

  • 04 Keep learning

    Iterate on new ideas

    We turn the reports into the next experiment: collecting what the evaluation already showed, deciding what to do next, and searching the literature for ideas that build on what you know and what is still worth exploring.

    • Next steps drawn from the reports and the evaluation

    • Literature research on what you know and could explore

    Browser window showing a housing project home page with an experiment backlog and ideas table

Benchmark

A working model, sooner.

Same model, same harness, same data. The skills change how long it takes to reach a working predictive model.

Time to a result

DeepSeek V4 Flash in OpenCode. Six open-data tasks. One point is one run.

Time to a working model averages 65.6 minutes with a plain coding loop and 36.9 minutes with the skills, across 18 runs each. Time to a result at least as good as a hand-written baseline averages 72.3 minutes with a plain coding loop, across 18 runs, and 48.9 minutes with the skills, across 17 runs. Points are colored Dataset 1 through Dataset 6.

  • Mean rounds

    1.28 / 1.67

    Rounds to a working model, with the skills first.

  • Round timeouts

    6 / 28

    Rounds that hit the time limit, with the skills first.

  • Final quality

    Level

    17 of 18 runs with the skills match a hand-written baseline, and 18 of 18 do without. The run that falls short is drinking-water.

Scale with Skore

Start locally.
Scale when you need it.

Your team needs more velocity

In the enterprise, velocity matters. Our skills bring methodological backing and statistical thinking. Skore is the platform that scales data science experiments: a remote agent on your LLM provider, remote compute on your infrastructure, remote storage so no result is lost, and a Skore UI where data scientists investigate the findings.

Discover Skore
  • 01

    Remote agent

    Connect your agent to the LLM provider you already use.

    02

    Remote compute

    Scale experiments onto your infrastructure.