Data Cleaning in Machine Learning Starts With Rules You Approve

What the agent can do

• Filesread_file, write_file, edit_file, grep and glob work inside the chat's working folder, and a counter shows files created, edited and read.
• Shellruns your Python command on a cleaning script and reports the real exit code; dangerous commands are blocked.
• Webweb_search and http_get read official documentation pages, for example when a rule needs a detail of the standard library.
• Screenshots and imagesocr reads text from a screenshot of a table, and write_file can save that text for a script to tidy.
• Orchestrationtask runs sub-agents in parallel, for example one per input file; run_on_model sends a hard step to a bigger model; skills hold reusable playbooks.
• Approval modesAuto-run acts at once, Ask first pauses before actions that change something, and Plan only writes a plan and runs nothing.
• Cronruns a prompt on a schedule, so a saved cleaning script can be re-run when a new export arrives.
• Modelsconnect a cloud provider through a standard API endpoint, a model running locally on your PC, or the built-in model; the Test button checks an endpoint.

Two walkthroughs for model preparation

Fill missing ages without leaking

Data preprocessing in machine learning adds steps after cleaning, such as encoding categories and scaling numbers, and every statistic must come from the training part only. Say you want the missing ages in ratings.csv filled with a median. Fredit Flow uses grep to find lines with empty fields, write_file to save prepare.py and Shell to run it with your own Python command. The script writes train_clean.csv and test_clean.csv, with the median calculated on the training rows and then applied to the test rows. Your ratings.csv stays untouched. In Ask first mode the card pauses before the write, so you read the diff, compare the row counts the script prints and spot-check 5 rows in each new file. A pipeline to clean messy real world data can start as this one script.

Check ranges and impossible values

Impossible values, such as a negative age or a negative price, hide in large files. If you are spending most of your time cleaning data for ML, write each rule down once and let the script apply it. Handling missing values, duplicates and outliers needs a separate rule for each, written down before the script runs. Fredit Flow uses read_file to check the column names, edit_file to add a range check to the script and Shell to run it, reporting the exit code and a count of rows outside the range. Rows that fail go to a review file. Nothing is deleted. Plan only mode lists these checks first. Read the review file line by line before you change a rule.