Machine learning and deep learning, without your data leaving your computer.

From the business goal to the published model in a guided six-step flow. Train with scikit-learn, PyTorch or Transformers, specialize local LLMs and answer from your documents. The agent helps you design and read results; the real data is only seen by you and your models.

Yuli Code — bajas-clientesML StudioNo network
  1. 1Home
  2. 2Data
  3. 3Design
  4. 4Train
  5. 5Evaluate
  6. 6Publish
BestF1 0,872
Pace · 38,4 steps/sTraining in progress
Finishes in4 min 12 s
Epoch 18 of 24 Loss (training) Loss (validation)
Validation metrics
  • F10,872
  • AUC0,931
  • Precision0,858
  • Recall0,887

Acceptance criterion: F1 ≥ 0.85 · Passes

clientes.csv · v4 · 48,210 rows
  • antiguedad_mesesint
  • plancat
  • email••••••
  • telefono••••••
  • rfc••••••
  • bajalabel

3 masked columns

Plan validated: 0 warnings

Offline worker · project .venv environment

Stratified split 70 / 15 / 15

Checkpoint saved to models/churn-gbm/v3

models/churn-gbm/v3GPU · 8,1 / 12 GBPython 3.12 · .venv

From idea to model in six steps

The numbered tabs guide you along. You can go back to any step.

  1. 1

    Home

    You describe the goal in your own words and Yuli translates it into ML: problem type, metric and strategy.

  2. 2

    Data

    You import and profile: nulls, balance, correlations, masked sensitive data and immutable versions.

  3. 3

    Design

    Yuli proposes the strategy and answers eight questions, from what it learns to how we’ll know it improved.

  4. 4

    Train

    Live learning curve, pause, resume, hyperparameter search and an overfitting warning.

  5. 5

    Evaluate

    Independent evaluation on data training never saw: robustness, fairness and champion versus challenger.

  6. 6

    Publish

    ONNX with parity, quantization, integrity tests and a local service on 127.0.0.1.

Everything you need to train for real

  • Deep learning

    Layer-by-layer network designer, your own tokenizers, GPU diagnostics, memory estimation and DDP.

  • LLM Studio

    Specialize local language models with SFT, DPO, ORPO, KTO and QLoRA, and quantize them behind a gate.

  • Knowledge (RAG)

    Answer from your documents: a local index, citations to the source and nothing sent to the internet.

  • Experiments

    Compare training runs, parameters and metrics; every model version is recorded on disk.

  • Approach advisor

    Tells you when NOT to use ML: if a rule is enough, if data is missing or if an LLM with RAG is the better fit.

  • With your agent

    Claude Code, Codex or Kiro operate ML Studio with MCP tools, and you see every step live.

Your data boundary

ML Studio applies the same rules that protect your code, now for your datasets.

  • The worker has no networkIt trains without internet access by default (allowNetwork: false).
  • The agent sees profiles, not rowsStatistics, schemas and metrics; anything sensitive is masked.
  • You decide what’s irreversiblePromoting a model to production always waits for a person.

Train your first model today

Create your account, open ML Studio with Ctrl+Alt+4 and pick a template or describe your problem.