Understand what went wrong.
Bring a small machine-learning project and its run results. Explore plausible explanations, the evidence behind them, and a next experiment.
Start an investigation
Upload a ZIP with Python code, config and run logs. Files are read as text, never executed. Selected contents are sent to Nebius for analysis. Remove private data and secrets before upload; automated checks are incomplete.
Demo access
Use the demo password, never a Nebius or Tavily API key. It is kept only in this page while open.
How it works
- Finds relevant files and metric lines.
- Uses Nemotron to build evidence-backed hypotheses.
- Suggests a controlled follow-up experiment.
- Uses Tavily to retrieve official documentation for the investigation when configured.
Model-generated explanations can be wrong. Check the cited evidence and avoid uploading secrets.
What the evidence suggests
Computed findings
Calculated from reported values in JSON; the run itself is not independently verified.
Citation checks
Checks whether lines were sent to the model and flags blank or punctuation-only citations. It cannot verify whether the citation supports the claim.
Numbers to review
Limited check for numbers not directly found in a single cited span. Derived gaps and ambiguous citations are skipped. A value may be rounded or expressed differently. No warning does not mean the answer is correct.