AutoML without uploading raw CSV data

The problem with upload-first AutoML

Upload-first AutoML asks users to move raw training data into a hosted system before they know whether the dataset is useful. That creates friction. A CSV may contain customer records, financial rows, operational details, patient-like attributes, employee information, or proprietary business metrics. Even when the dataset is ultimately safe to process externally, the early exploration phase is often full of uncertainty.

Teams may ask basic questions first: is the target usable, are there enough rows, are categories meaningful, are missing values concentrated in a time period, is the data sorted, and are there columns that leak the answer? It is reasonable to answer those questions with less data movement. AutoML without uploading raw CSV data is about shifting early exploration closer to the user's device and delaying heavier infrastructure decisions until the dataset has earned them.

How browser-local training changes the workflow

MLdeck brings the training workflow to the browser. Instead of sending the CSV to a cloud training job, the app parses and profiles the file locally, lets the user select features and a target, and fits candidate tabular models using browser-executed tooling. During normal browser training flows, raw CSV data is not uploaded to MLdeck servers.

This does not mean every part of the product is disconnected from backend services. Account, app, support, billing, security, and control-plane features may use backend infrastructure. Optional AI Copilot and Optional AI Advisory Notes can send sanitized metadata after session consent, including column names, target name, metrics, preprocessing steps, validation/export status, and the user's question. Those optional AI features do not send raw CSV rows, row samples, uploaded file contents, model binaries, package binaries, prediction curves, or large artifacts. The important distinction is that normal CSV training does not require raw CSV cloud upload for model fitting. One network dependency exists at training start: the browser ML runtime may be downloaded from a third-party CDN. This is a code download — raw CSV rows and file bytes are not sent to it. See the privacy-first page for details.

What MLdeck does locally

MLdeck's browser workflow covers the practical pieces of early tabular modeling. It profiles columns, estimates data types, shows missingness and cardinality, supports feature inclusion decisions, helps identify targets, fits classification or regression candidates, displays leaderboard evidence, and prepares ONNX-oriented export packages. Users can inspect warnings and decide whether the dataset is ready for more serious validation.

The local approach also helps education. Students can open a CSV and see how preprocessing affects model evidence without installing Python, configuring notebooks, or receiving cloud credentials. Analysts can test a hypothesis before asking for engineering help. Developers can generate ONNX-oriented export packages for parity checks before designing a serving system.

What still requires caution

Privacy-first architecture does not remove user responsibility. Your browser, extensions, operating system, local storage policies, downloaded reports, exported models, and shared files all matter. If a CSV is sensitive, derived artifacts may also be sensitive. A model can encode patterns from source data, and a PDF report may include metrics or field names that should be handled carefully.

MLdeck is a browser-local AutoML product for privacy-first CSV modeling. It helps users profile data, train models locally, review warnings, and prepare reports while keeping raw CSV data in the browser during the normal training workflow. Users working with regulated or high-impact data should also follow internal review procedures and legal guidance.

Practical checks include using a trusted browser profile, closing unneeded extensions, confirming the file is allowed for local analysis, reviewing data quality for machine learning, and deciding where exported artifacts will be stored after download. The browser-local training flow is one part of a broader data handling process.

Exporting results without locking data into a cloud platform

One benefit of a browser-first workflow is that the result is not tied to a proprietary hosted endpoint. MLdeck can generate ONNX-oriented export packages designed for ONNX Runtime testing with schema, manifest, and parity-review metadata. It can also produce Docker packages for runtime testing and PDF reports for review. These artifacts support a workflow where early exploration stays local and later validation can happen in an environment chosen by the user.

Exports are testable artifacts. Teams can review input schema, preprocessing metadata, validation samples, behavior across representative rows, and monitoring requirements before moving toward important systems.

AutoML without uploading data FAQ