Documentation / ML Studio

ML privacy and sandbox

ML Studio's data boundary. What stays on your computer, what agents see, who unmasks and what happens before training elsewhere.

On this page

In ML Studio there are two different audiences and Yuli keeps them apart: you and your local models, which work with the real data, and external agents, which only see what passes through the data boundary.

Your code lives on your computer; with a local model nothing leaves; with cloud models, whatever is sent first goes through the masking of secrets and personal data.

What stays on your computer

All of ML Studio runs locally:

ActivityWhere it happens
Importing and profiling dataOn your computer.
Training and evaluatingIn the ML worker on your computer.
Serving the modelOn 127.0.0.1, listening only on your machine.
Local LLM and RAG (Knowledge)On your computer, with built-in offline embeddings.

There are no Yuli servers between you and your AI. The only exception is the one you choose explicitly: training on a remote machine (see below).

The data boundary

External agents (Claude Code, Codex, ACP agents or a cloud model behind the chat) send what they read to their provider. That's why everything ML Studio returns to them goes through a data boundary:

  • Sensitive columns go out as •••.
  • Emails, cards, phone numbers, CURP/RFC and IP addresses are removed from free text.
  • Aggregates do get through: profiles, metrics and feature importance. An agent can reason about “which variables matter most” without seeing a single customer.

Example of what an agent sees when it asks for a sample:

text
cliente_id  email   curp    saldo_prom  pedidos_pagados  baja
•••         •••     •••     1 840.50    12               0
•••         •••     •••       312.00     2               1

With a cloud model, in addition, whatever is sent to the provider goes through the egress filter: secrets are replaced with a yuli:secret#N placeholder and personal data with a yuli:pii#N one, with a per-session HMAC tag. They are only restored when writing to disk. See Secrets and privacy.

Sensitivity detector

In Data › Sensitivity each column gets a probability, a category and a suggested masking. The detector combines the column name, patterns and check-digit validators:

ValidatorDetects
CardsPayment card numbers.
IBANInternational bank accounts.
CLABEMexican bank accounts.
CPFBrazilian tax ID.
CURP / RFCMexican personal and tax ID.

Who unmasks

Only a person can mark a column as not sensitive. An agent can never unmask data for itself, not even inside an agent team.

What you decide:

  • Mark as sensitive a column that Yuli didn't detect.
  • Mark as not sensitive one that was detected by mistake.
  • Create a masked copy to share the dataset.

Offline worker by default

The process that trains (the ML worker) has no network access by default:

MeasureEffect
HF_HUB_OFFLINEHugging Face libraries don't download anything.
PIP_NO_INDEXpip doesn't install packages from the internet.
bwrap --unshare-net (Linux)The process runs without a network interface.

That way, a training script or a dependency can't send your data anywhere. The network is only opened if the project allows it explicitly:

yaml
# yuli-ml.yaml
policies:
  allowNetwork: true   # for example, to download a base from the catalog

Turn on allowNetwork only while you need it (to download a base or a dependency) and set it back to false afterwards.

Before training elsewhere

Training on a remote machine (SSH or cloud) is the only way ML data leaves your computer, and it never happens without a warning. Before launching, Yuli shows:

“Data leaves for host: the dataset and the code the job reads are uploaded. When it finishes, Yuli brings back the model, verifies every file and deletes the data there (unless the machine keeps it).”

0 of 4 done