How Karg Works
Two ways to clean your data: run Karg yourself, or let our team do it. Open a card to see each step.
Ready to try it on your own data? Run the platform yourself, or let our team do the cleaning for you.
Two ways to clean your data: run Karg yourself, or let our team do it. Open a card to see each step.
Ready to try it on your own data? Run the platform yourself, or let our team do the cleaning for you.
Six steps from your original file to a result you can use.
Drop in a CSV, TSV, Excel, JSON, NDJSON, Parquet, or XML file. Karg stores it, opens a draft run, and takes you straight to setup.
If you saved the cleaning plan from an earlier run as a template, you can reuse it here in one click.
Karg reads the file first, then asks only what the file leaves open: how dates should be written, what to do with incomplete emails, whether blanks may be filled, and how much it may apply on its own.
Every question says why it is being asked, and Not sure is always an answer. You finish with a plan written in plain language. Nothing in your file changes until you start cleaning.
Cleaning runs in front of you, stage by stage, with a Stop button the whole time. Stopping writes nothing.
Follow the progress as Karg checks your file, applies confident changes, and sets aside suggestions that need your decision.
The changes Karg was not confident enough to apply are waiting in a queue. Each one shows the source value beside the suggestion and the reasoning behind it.
Approve or reject with A and R, and reopen any decision before you finalize. Suggestions you do not approve are never applied to the final file.
Open the cleaned file in the spreadsheet editor to inspect the data and make any final adjustments before downloading.
Review your edits, then commit them to the run so the result includes the changes you chose to make.
Download the cleaned file in its original format, or as CSV, Excel, or JSON.
Suggestions you did not approve are never applied to it.
With Karg Applied, you send the files and our team runs them through the same engine, then sends them back clean.
Send a short request with what the data is, where it comes from and what it is for. We reply to set up a call and agree on scope, deliverables and how files are shared.
Our team runs your files through the same Karg engine the platform uses: rules, specialized machine learning models and AI database agents, with every uncertain change checked by a person.
You get the cleaned files back, with the exceptions we could not resolve and a summary of what changed. Nothing new to install or learn.
Karg profiles your file, applies the changes it is sure of in three layers (rules, specialized machine learning models and AI database agents) and sets aside the ones it is not sure of. Our team decides those with you, then sends back the clean data and a summary of what changed.
By agreement, files can come by email or a share link, with no account to create. How your files are handled.
A cleaned file, and a record you can check it against.
In the format you uploaded, or as CSV, Excel, or JSON. It is a separate file named after your upload; Karg works on a copy and does not overwrite what you sent.
One entry for each transformation Karg applied: the column, the action, the reason, the confidence, and whether a rule, a model, or AI made the call.
The transformation plan and the setup you agreed to at the start, as downloadable files, so the result can be traced back to your answers. A finished plan can be saved as a template for the next file.
In CSV and Excel downloads, any cell a spreadsheet would run as a formula is escaped, so opening the file cannot execute something a stranger typed into your data.
A second file that checks the run itself: how your file was read, whether every row was preserved, and a measured comparison of output to input that flags any change no transformation accounts for.
Download data package pairs the cleaned CSV with a lineage summary: the run's signature, when it finished, records cleaned, transformations applied, columns touched, and the lineage steps.
The log is plain JSON. Read it, hand it to an auditor, or load it into whatever you already use. Each entry says which column was touched, what was done, and how sure Karg was.
The log is organized by transformation and column. It does not list a row number for every changed cell.
A file is available before you finish review; it contains only the changes Karg applied on its own. Finalizing review produces the file with your approved changes included.
Links to your files are signed and expire after one hour by default.
Plain answers to the questions people ask before they upload real data.
Karg works on a copy and writes a separate cleaned file. Your original is preserved during the job, then deleted along with the rest of the run.
Only changes at or above your confidence threshold. The default is 0.85 and you can move it in settings, or choose at setup to have Karg flag everything and apply nothing. Removing rows is held to a higher bar. Anything below the line waits for your review.
Rules first, specialized machine learning models next, frontier LLMs only for the most ambiguous transformations (if enabled).
A profile of each column the rules could not settle: its name and type, counts, its most common values, and a small sample of values. Never your entire file. Requests go only to zero-data-retention endpoints with data collection denied, so your data never reaches an inference provider that trains on prompts.
Whole files are never used for training. The only training data is review labels, and only if your organization opts in.
No other organization. Every request is checked against your organization, the database enforces row-level security underneath that, and traffic is encrypted in transit.
Your files and results are deleted from active systems within 72 hours of the job completing or failing. Your organization can shorten that window but not extend it, and you can delete a run yourself at any time. Backups can persist for up to 7 more days.
Yes. Approve, reject, and reopen any decision until you finalize. Suggestions you did not approve are never applied to your download.
Karg does not hold SOC 2, ISO 27001, or HIPAA certification. Do not upload health records, card numbers, Social Security or government ID numbers, or bank details.
The full policy, with the detail behind each of these answers, is on the Security page. For what is under the hood, see the Product page.