Skip to content
NautAI

Everything between raw data and a better model.

Seven parts of one versioned workflow. Use all of them, or start with the one that hurts most and connect the rest later.

Collect

Bring data in from wherever it already lives.

Connect object storage, warehouses, databases and event streams, or upload files directly. Imports are hashed row by row, so re-importing the same export only adds what changed.

  • Object storage, Postgres, MySQL, warehouses and Kafka-compatible streams
  • Text, images, audio, conversations and tabular rows
  • Scheduled syncs with schema checks before anything lands
sources / support-tickets
  • s3://acme-support/exports

    Object storage

    1.18M
  • postgres / tickets_prod

    Database

    live
  • events.chat_transcripts

    Stream

    4.1k / day
  • uploads / handwritten-forms.zip

    Upload

    pending

Clean

Cleaning as a pipeline, not a notebook.

Describe your cleaning steps in a short YAML file. NautAI runs them on every import and records exactly which rows each step touched, so you can explain any row's history.

  • Exact and embedding-based near-duplicate detection
  • PII detection and redaction for emails, phones, names and card numbers
  • Quality scoring with your own filters or a model you choose
pipeline / clean.yaml run #128
Cleaning results
Near-duplicates24,117 rowsDropped
PII: emails, phones, card numbers8,902 rowsRedacted
Language not English3,440 rowsMoved to branch
Quality score below 0.41,206 rowsSent to review

Label

Review queues your annotators will actually like.

Split work into queues with written guidelines, pre-labels from any model and consensus rules. Bring your own team or annotators; Enterprise plans can add a managed workforce.

  • Model-assisted pre-labeling with confidence thresholds
  • Consensus, adjudication and per-reviewer agreement stats
  • Guideline versions linked to the labels they produced
queue / refund-vs-billing (guideline v3)

“I upgraded by mistake yesterday, can you reverse it and put me back on the old plan?”

Pre-label from model: refund (0.71)

billingrefundplan changecancellation
Agreement with 2 other reviewers: 2 of 3418 / 2,130

Synthesize

Synthetic data with a paper trail.

Spot classes and edge cases your dataset barely covers, generate examples seeded from real rows, and send them through the same review as everything else.

  • Coverage reports by label, language, length or any field
  • Generation seeded from real rows, with your model or ours
  • Every synthetic row is tagged, filterable and attributable
synthesize / fill under-represented intents
billing
34%
bug report
27%
refund
22%
cancellation
18%
accessibility
12%
Real rows Synthetic, reviewedSeeded from 1,840 real examples

Version

Commits, branches and merge requests for rows.

Datasets in NautAI work the way your code does. Experiment on a branch, review the diff, merge when evals pass, and tag the exact version each model trained on.

  • Row-level diffs between any two versions
  • Protected branches and required reviewers
  • Tags and lineage that link each model to its training data
naut diff v2.3..v2.4 --sample 6
ChangeRowtextlabel
r-20418I was charged twice for my plan this monthbilling
−r-20419Can I get my money back for the annual plan?billing
+r-20419Can I get my money back for the annual plan?refund
−r-20422pls call me on +1 415 555 0182callback
+r-20422pls call me on [PHONE]callback
+r-31007The export button does nothing on Safaribug

2,130 relabeled, 8,902 redacted, 12,408 added

Evaluate

Evals that run every time the data changes.

Attach eval suites to a dataset and NautAI runs them on each commit against the model you choose. When a score moves, you can trace it to the commit and the rows behind it.

  • Bring your own metrics or use built-in classification, retrieval and LLM-judge evals
  • Score history across versions and branches
  • Merge checks that block changes which lower a score
evals / intent-accuracy across versions
0.800.850.900.95v1.8v1.9v2.0v2.1v2.2v2.3v2.4relabeled refund intents

Feedback loops

Production tells you what to label next.

Send thumbs-downs, escalations and low-confidence predictions from your app with one SDK call. NautAI groups them, routes them to review and stages them for the next version.

  • SDK and webhook ingestion from any app
  • Automatic grouping of similar failures
  • Weekly release trains for dataset versions
feedback / production → review queue
  • Where's my invoice for March?

    Thumbs down

    billing
  • cancel but keep my data pls

    Agent escalated

    cancellation
  • Does the API support webhooks?

    Low confidence (0.38)

    unlabeled

312 rows captured this week, queued for support-tickets v2.5

Try it on a dataset you already have.