DataReady

• 

Inspiration

I wanted to make AI-assisted data preparation easier to verify and safer to use in everyday business work.

A real hotel directory from my restaurant outreach project made the challenges concrete: introductory spreadsheet rows, public business contacts, duplicate records, and blank fields that may be perfectly acceptable. A tool needs business context before treating every finding as something to fix.

DataReady brings together AI suggestions, independent checks, and human approval.

What it does

DataReady accepts CSV, flat JSON, and Excel files. It identifies worksheet layouts and header rows, checks the selected table, and presents findings for review.

Users can declare supported public business contact columns and optional blank fields, then recheck the unchanged data. These declarations record the reviewer’s judgment; other critical findings remain blocking.

For eligible inputs, NVIDIA Nemotron proposes repair candidates through Nebius Token Factory. DataReady checks each proposal against predefined rules, tests permitted changes on temporary copies, and audits the results again. A recommendation depends on measured improvement and regression checks.

The user explicitly approves a separate repaired CSV. The original remains unchanged, and a downloadable JSON report records audit findings, review decisions, and verification evidence.

A built-in sample lets visitors try the workflow themselves.

How we built it

I built the application with Python, FastAPI, Pydantic, pandas, openpyxl, and a browser-based upload interface. It is deployed on Google Cloud Run.

The live reasoning provider is NVIDIA Nemotron 3 Nano 30B A3B, accessed through the Nebius Token Factory inference API. The model receives minimized audit evidence with aliased column names; raw dataset rows are not sent to it.

The repair engine supports controlled operations such as exact-duplicate removal, whitespace trimming, and missing-marker normalization. The model proposes structured actions; it cannot run arbitrary generated code or authorize an export.

SHA-256 fingerprints connect the source, repair trial, and downloaded result. Secret Manager stores runtime credentials, while Firestore supports server-enforced inference quotas.

During the hackathon, I substantially extended an earlier prototype with the NVIDIA/Nebius integration, competing repair candidates, independent trials, the upload interface, Excel table detection, scoped human review, verified export, and public deployment controls.

Challenges we ran into

Real spreadsheets required more than simply reading the first row as a header. DataReady needed table detection, worksheet selection, and a way for users to correct the detected layout.

The hotel workbook also exposed a longitude value incorrectly flagged as possible personal information. This led to coordinate-range validation and more careful handling of business context.

Another challenge was separating review from repair. Confirming that a blank field is optional should record a decision without inventing a value. Likewise, a plausible AI proposal should not become a recommendation unless its measured result supports it.

Accomplishments that we're proud of

• Deployed a working public application with live NVIDIA inference through Nebius Token Factory.
• Demonstrated human review on a real hotel workbook without changing its values.
• Compared two AI repair candidates in the recorded sample demonstration.
• Removed one exact duplicate, reducing four rows to three while preserving the missing category.
• Verified that an approved downloaded CSV matched its independently tested result.
• Passed 393 automated tests and repository lint checks.

What we learned

Business context matters as much as detecting a pattern. Blank fields can be intentional, and a contact-information warning does not establish whether a column contains private or public information.

We also learned that valid AI output is only the beginning. Useful repair requires checking permissions, measuring the outcome, preserving unresolved findings, and making approval understandable.

A quality score is a heuristic—not proof that data is correct or ready for every business purpose.

What's next for DataReady

Next, I want to make review decisions and score explanations clearer, support user-defined data requirements, and evaluate the workflow across more representative datasets.

Further work includes benchmarking model latency, cost, and proposal quality; expanding tested repair operations; and adding authenticated workspaces and stronger isolation for private business use.

The guiding principle remains: AI proposes, independent checks verify, and the user approves.

Built With

• css
• fastapi
• git
• github
• github-actions
• google-artifact-registry
• google-cloud
• google-cloud-build
• google-cloud-firestore
• google-cloud-run
• google-secret-manager
• hmac
• html
• javascript
• nebius-token-factory
• nvidia-nemotron
• openpyxl
• pandas
• pydantic
• pytest
• python
• python-multipart
• ruff
• sha-256
• uvicorn