Your training data, under version control.
NautAI is where AI teams collect, clean, label and evaluate datasets, then feed what happens in production back into the next version.
$ naut push ./tickets -m "Add October escalations"Hashed 12,719 rows, skipped 311 duplicates✓ Eval suite intent-accuracy: 0.94 (+0.02)✓ Published support-tickets v2.4
Data teams at Halyard Labs, Kestrel Speech, Orrery Health, Lumen Robotics, Tidewater Legal AI and Parallax Search keep their datasets in NautAI.
A history for every row.
Every change to a dataset is a commit: who changed it, which rows moved, and what it did to your evals. Branch to try a new labeling guideline, open a merge request for review, and roll back a bad import with one command.
- Row-level diffs between any two versions
- Branches, merge requests and required reviews
- Tags that pin the exact data a model was trained on
Add October escalation tickets
a3f91c2 by data-engv2.4
+12,408 −311eval 0.94 +0.02Relabel refund intents with guideline v3
7be04d9 by labeling
+0 ~2,130eval 0.92 +0.03Redact emails and phone numbers
c18e7a0 by pipeline
+0 ~8,902eval 0.89 0.00Drop near-duplicates (cosine ≥ 0.97)
52d6b3e by pipeline
+0 −24,117eval 0.89 +0.01Import from support warehouse
e09a114 by data-eng
+1.18M −0eval 0.88 baseline
From raw data to a better model, and back again.
Most teams build this loop out of scripts, spreadsheets and a labeling tool someone wrote during a hackathon. NautAI runs it as one versioned pipeline.
- 1
Collect
Connect buckets, databases, event streams and production logs. Every import becomes a commit.
- 2
Clean
Drop near-duplicates, redact PII and filter by quality score, as a pipeline you can rerun.
- 3
Label
Send rows to review queues with model pre-labels, written guidelines and consensus checks.
- 4
Evaluate
Run eval suites against each version and see which change moved the score.
- 5
Improve
Route production failures back into review, then ship them in the next version.
Then the course repeats, starting from what the last version got wrong.
| Change | Row | text | label |
|---|---|---|---|
| r-20418 | I was charged twice for my plan this month | billing | |
| − | r-20419 | Can I get my money back for the annual plan? | billing |
| + | r-20419 | Can I get my money back for the annual plan? | refund |
| − | r-20422 | pls call me on +1 415 555 0182 | callback |
| + | r-20422 | pls call me on [PHONE] | callback |
| + | r-31007 | The export button does nothing on Safari | bug |
2,130 relabeled, 8,902 redacted, 12,408 added
Review data changes like code.
naut diff shows exactly which rows were added, relabeled or redacted. Reviewers comment on rows, request changes and approve, so nothing reaches training without a second pair of eyes.
Fill the gaps with synthetic data you can trace.
Find classes your model rarely sees, generate examples seeded from real rows, and review them before they merge. Synthetic rows are tagged forever, so you can always filter them out or measure what they contributed.
Know which version made the model better.
Attach eval suites to a dataset and they run on every commit. When a score moves, NautAI points to the change that moved it, down to the rows.
“We used to find out a relabeling job had gone wrong two weeks later, from a regression in production. Now it shows up in the merge request.”
“Our S3 folder called final_v2_really_final is gone. Every model now points at a tagged dataset version.”
“Production feedback goes straight into a review queue. We ship a better dataset every week instead of every quarter.”
What you can stop maintaining.
Data infrastructure takes months to build and never feels finished. NautAI replaces the pieces most AI teams end up writing themselves.
Dedupe and PII scripts that only one person understandsRerunnable cleaning pipelinesA labeling tool built during a hackathonReview queues with guidelines and consensusA spreadsheet of which CSV trained which modelTagged versions linked to each training runNotebooks that rerun evals by handEval suites that run on every commit
Chart your first dataset this afternoon.
The free plan includes three private datasets, unlimited public ones and 5,000 synthetic rows a month. No card needed.
$ pip install nautai $ naut init