Browser-Local AutoML for Private CSV Model Training

What browser-local AutoML means

Browser-local AutoML moves the first modeling loop into the browser. Instead of starting with cloud training infrastructure or a Python notebook setup, users can upload a CSV, inspect columns, choose a target, configure preprocessing, train candidate models, compare metrics, and review warnings in one local workflow.

This approach is designed for early-stage tabular model exploration and validation-aware review. It helps users compare models, inspect metrics, review data-quality warnings, generate local Advisory Playbooks, and export ONNX-oriented packages for further testing. Production use requires independent validation, parity checks, and environment-specific testing.

How MLdeck keeps raw CSV workflows local

Core CSV profiling, preprocessing, browser-side model training, model comparison, and local Advisory Playbook generation are designed to run in the browser during normal workflows. Raw CSV rows do not need to be uploaded to a cloud training job just to explore whether a dataset has signal.

Privacy-first does not mean risk-free. Users remain responsible for local device security, browser extensions, downloaded artifacts, internal data policies, and any external systems used after export. Derived artifacts such as reports, schemas, and model files should be handled carefully when source data is sensitive.

Validation-aware model review

MLdeck surfaces baseline-aware model comparison, data-quality warnings, leakage-risk review, feature importance, and evidence labels so users can interpret results with context. Exploratory and prequential metrics are useful for early review, but they are not independent holdout evidence.

If the best model does not beat the baseline, it should not be treated as a deployment candidate. Improve data quality, feature selection, target definition, or model quality before deployment validation. Strict validation should be run before deployment decisions or other important reliance on model outputs.

Local Advisory Playbooks and optional AI Copilot

MLdeck’s Advisory Playbook has two layers. Sections 1-10 are generated locally in the browser from the current run metadata. Optional AI Advisory Notes can be added only when the user enables Optional AI Features and grants metadata-sharing consent.

Optional AI Copilot features can explain metrics, validation limits, feature importance, export readiness, and next experiments. AI guidance uses sanitized metadata only after consent and does not inspect raw rows or model binaries. Metadata that may be sent includes dataset name, column names, target name, model names, metrics, preprocessing steps, feature-importance names, validation/export status, and the user’s Copilot question. Raw CSV rows, row samples, uploaded file contents, ONNX/PKL/model binaries, Docker/package binaries, prediction curves, and large artifacts are not sent by those optional AI features.

Playbooks and Copilot answers are snapshots of a run context. If the dataset, target, features, model, validation status, or export status changes, older Playbooks or Copilot answers may be marked as previous-run or stale.

ONNX-oriented export packages for external validation and testing

MLdeck can generate ONNX-oriented export packages from supported browser-local training workflows. These packages are intended for external validation, schema review, parity testing, and runtime experiments. They are not a shortcut around validation.

ONNX artifact availability means an artifact exists. It does not prove parity, inference fidelity, or deployment readiness. A parity check, representative external validation data, and environment-specific testing are still required before relying on exported artifacts.

Limits and responsible use

MLdeck is an MVP / early beta product. Browser CPU and memory limits apply, some models or exports may be skipped, and large or high-dimensional datasets may require dedicated infrastructure. Use MLdeck to learn quickly, reduce raw-data movement during early work, and prepare validation artifacts. Use strict validation, domain review, and external testing before relying on outputs in important settings.