Sushi ML Studio: No-Code Machine Learning with scikit-learn

Turn a spreadsheet into predictions you can trust, without writing code, hiring a data scientist, or sending your data to anyone.

Sushi ML Studio is a web application that runs on an EC2 instance in your own AWS account. Open it in your browser, upload a file, and within minutes you have a trained model, an honest account of how good it is, and predictions for new rows.

HOW IT WORKS

1. Load your data. Upload a CSV or Excel (.xlsx) file up to 50 MB, or start from one of three example datasets. The Studio reads the file, detects separators and encodings, and profiles every column: its type, blank cells, typical values, and warnings about columns that would mislead a model, such as IDs, dates, free text and duplicates of each other.
2. Choose what to predict. Tick the column you want answers about. The Studio works out whether that means sorting rows into groups (classification, for example churned or stayed) or predicting a number (regression, for example a sale price), and you can override it.
3. Train. Seven scikit-learn models are trained and compared for you, including gradient boosting, random forest, and linear models. An optional search tries better settings, and then tells you whether it actually helped.
4. Read the results. A plain-language verdict says how good the model is, the predictions and metrics are charted, and a ranking shows which columns drive the prediction.
5. Use it. Upload new rows, as CSV or Excel, and download them as a CSV file with a prediction column added.

RESULTS YOU CAN TRUST

• Every score is measured on rows the model never saw during training, using cross-validation on smaller files and a held-back test set on larger ones.
• Every score sits next to what a simple guess would have achieved, so a 95% accuracy on data that is 95% one answer is called out, not celebrated.
• When the models have not learned anything useful, the Studio says so.
• When several models perform the same within noise, it says they are tied and recommends the simplest one, instead of crowning a winner by a hair.
• Column importance is measured on held-out rows, and columns that duplicate each other are named, so the ranking does not mislead.

YOU KEEP WHAT YOU MAKE

• Download the trained model as a standard scikit-learn file that loads with joblib, with no dependency on this product.
• Download the complete Python training code, with the random seeds used, so running it reproduces the same numbers.
• The instance also includes JupyterLab and a Python environment with scikit-learn, pandas, NumPy, SciPy, matplotlib and seaborn, for when you want to continue in code. The Studio runs from its own separate environment, so changing packages cannot break it.

PRIVATE BY DESIGN

• Runs entirely on your instance, in your account and your chosen region. Uploaded data is held in memory and never written to disk or sent anywhere.
• No accounts, no telemetry, no licence server, no calls home.
• The Studio is password protected from the first boot. The initial password is your instance ID, which only people with access to your EC2 console can see, and you can change it from the Studio at any time.

GOOD TO KNOW

• Works with tabular data: one row per example, one column per attribute. It does not do time-series forecasting, images, or text analysis.
• Download your model before you close the browser tab. The Studio does not keep projects between sessions, and warns you before anything unsaved would be lost.
• Recommended instance type: t3.medium or larger.

Built by Software Sushi, a team that designs custom machine learning and AI systems. Support is by email on business days.