Documentation / ML Studio

What is ML Studio

Yuli's machine learning space. Data, experiments, training, models and deployment in a guided six-step flow, all on your computer.

On this page

ML Studio is Yuli's machine learning space: data, experiments, training, models and deployment. You recognize it by its violet tone in the left rail and you open it with Ctrl+Alt+4 (⌘+Alt+4 on macOS).

Everything happens on your computer: importing data, training, evaluating, serving the model on 127.0.0.1, the local LLM and the knowledge base (RAG). The agent helps you design and read results, but the real data is only seen by you and your local models. See ML privacy and sandbox.

ML Studio trains in Python, in an environment of its own for each project. It requires Python 3.11 or later. Libraries are never installed without your permission, and they go into the project's environment, not the system.

Start here

The Home tab opens with “✦ Start here” and a question: “What do you want your model to learn?”. “Speak from your field. Yuli translates it into ML.” Describe the problem in your own words or pick an example:

BusinessOperationsText and language
Predict churnTime seriesSpecialize an LLM
Forecast salesDetect anomaliesClassify logs
Default riskRecommend productsClassify tickets
Segment customersEffect of a campaignDetect sensitive data

Press Let Yuli propose it ✦ and the agent translates your goal into a problem type, a metric and a strategy.

While a model trains, the home page shows the live status: Best (the best metric so far), Pace · n steps/s and Finishes in. If the validation loss starts to rise, it warns you: “The validation loss is rising from its minimum: possible overfitting.”

Create an ML project

The wizard asks “How do you want to start this model?”:

OptionWhen to choose it
From a templateTo learn or get going quickly with sample data.
With my own modelYou already have a script or a model and want to see it in ML Studio.
With the agentYou describe the problem and the agent prepares the project.
From scratchAn empty project that you configure yourself.

Starting frameworks: scikit-learn (classic), PyTorch (neural network) and Transformers (text). Templates: Tabular classification, Fraud detection, Neural network (PyTorch), Regression and Text classification.

Yuli creates the yuli-ml.yaml manifest at the project root. A project can have several models; the list lives in ml/use-cases.json.

yaml
# yuli-ml.yaml
apiVersion: yuli.ml/v1
project:
  name: bajas-clientes
  task: classification
runtime:
  python: "3.12"
  framework: sklearn
hardware:
  prefer: auto          # gpu, cpu, remote or auto
policies:
  maxConcurrentJobs: 1
  allowNetwork: false   # the ML worker has no network by default
  retainCheckpoints: 3
acceptance:
  minimum_f1: 0.85
  maximum_latency_ms: 200

The guided 6-step flow

The numbered tabs take you from the idea to the published model. You can go back to any step.

1 of 6

1 · Home

You describe the goal and Yuli proposes the approach. Here you also see the live status of whatever is training.

More tools

Besides the guided flow, ML Studio brings tools for those who want to go further:

ToolWhat for
Learning typesChoose between classification, regression, anomalies, series, recommendation, reinforcement…
Model explorerSee on the inside what the model learned.
Architecture designerBuild neural networks layer by layer. See Deep learning and GPU.
ExperimentsCompare training runs, parameters and metrics.
LLM Studio (“Language”)Specialize language models. See LLM Studio and RAG.
Feature LabCreate and test derived features.
Knowledge (RAG)Answer from your documents.
IntelligenceML agents and the Approach advisor.
OperationsPipelines, a gateway with canary, shadow and rollback, and your extensions.
Machine learning guideA complete guide, from the basics to the formulas.

When NOT to use ML

Not every problem needs a model. The Approach advisor analyzes your goal and your data and answers with one of these verdicts:

VerdictWhat it means
A model adds valueThere is a pattern, enough data and a clear metric. Go ahead.
A rule is enoughA simple condition solves the case just as well or better, and is easier to maintain.
No ML neededThe problem is deterministic: a query, a calculation or a validation.
LLM with RAGYou need to answer from your documents, not predict a column.
Not enough dataThere aren't enough examples, or enough labeled ones, to learn from yet.

Always start by comparing against the simplest possible baseline (the most frequent class, the mean). If the model doesn't clearly beat it on data it hasn't seen, you don't need it yet.

Working with the agent

Agents operate ML Studio with their own MCP tools: ml_project, ml_data, ml_tokenizer, ml_architecture, ml_training, ml_experiment, ml_evaluation, ml_inference, ml_model, ml_resources and ml_agents. In an agent team there are specific roles: data-engineer, ml-engineer, evaluator and mlops.

Ask it things like “train a model that predicts churn with this data” or “why does this model fail with new customers?”. You'll see every step live in ML Studio, and anything irreversible (promoting to production) is always decided by a person.