AI data cleansing for mid-market teams
Connect a warehouse or drop a CSV. Data Stew profiles every column, finds the issues, and proposes fixes with rationale and evidence — nothing touches your data until a person approves it.
Free tier includes 10 AI fix suggestions and a shareable quality report. No card required.
| Column | Current | Proposed | Confidence | Why |
|---|---|---|---|---|
| state | N.S.W. | NSW | 98% | Canonical form — 312 rows follow this pattern |
| segment | ∅ missing | Wholesale | 93% | 3 similar records from the same supplier agree |
| phone | 412345678 | +61 412 345 678 | 96% | Format rule learned from 97% of the column |
| abn | 51 824 753 5 | — abstained | needs you | Evidence insufficient; flagged for a person instead of guessed |
What it does
Detection tools tell you what’s broken and stop. Data Stew owns the fix: suggestion, approval, writeback, and the score that proves it worked.
01
Null rates, format distributions, near-duplicate categories, cross-column dependencies — computed as push-down SQL in your warehouse, or DuckDB for files. Rules are generated from what your data actually looks like, and every dataset gets an explainable quality score.
02
AI looks at similar records the way a person would, then proposes a value with its confidence, rationale, and the evidence rows it used. When the evidence is thin, it abstains — a wrong guess costs more trust than a blank.
Sydny → Sydney 97%
“14 of 15 records at this postcode read Sydney; the outlier matches no known suburb.”
03
A review grid built for throughput: approve, edit, or reject; bulk-accept everything above a confidence bar. Systematic fixes are approved once as a pattern — one decision can correct thousands of rows.
Pattern: canonicalize state namesapproved
Rows corrected by this decision312
04
Fixes land as a suggestion table and a script your team runs — or export a cleaned file. Who approved what, which model proposed it, under which prompt version: an append-only audit log built for your auditors.
How it works
1
Upload a CSV, or connect Postgres, MySQL, SQL Server, Snowflake, or BigQuery with a read-only role — the least-privilege setup is generated for you.
2
About a minute later: a score, a per-column issue breakdown, and sample AI fixes with their reasoning. Shareable with anyone, no account needed.
3
Work the review queue, bulk-accept the confident ones, and export the cleaned file or the fix script for your warehouse.
~60 sec
from CSV upload to quality report
10,000+
cells one approved pattern can fix
2
AI providers allowed to see cell data
Why teams switch
vs Legacy DQ suites
Six-month implementations, six-figure contracts, a services team on site.
Connect to first fix suggestions in under an hour, from ~$500 a month. No implementation project.
vs Detection-only tools
Dev-centric alerting: you learn what’s broken, and the tickets pile up.
We own the fix — suggestion, steward approval, writeback — operated by the business, not the backlog.
vs Black-box ML mastering
A model changes your records and asks you to trust it.
Every fix carries its rationale, evidence, and approval trail. AI proposes, your people dispose.
Pricing
Start free with one file and ten suggested fixes. Paid plans meter on records suggested, with credit top-ups when a big cleanup runs long.
$0 / month
See what’s wrong with one file, with the reasoning behind every fix.
$500 / month
One source, a steward and a reviewer, the full approve-and-apply loop.
$1500 / month
Several sources, a steward per domain, and the dashboards to run it.
Let's talk
EU residency, BYOK, SSO, and a contract your security team has read.
Upload one messy file and see the score, the issues, and ten proposed fixes — with the reasoning — in about a minute.