Our summary
An incident report can mix customer complaints, a timeline and unfinished investigations. Olga Stefaniuk shows how Box turns a PDF into readable text that keeps headings and tables. Jev then judges the incident's kind, severity and need for urgent escalation, rather than writing another summary for someone to interpret.
Ordinary Python code chooses the next step. An escalation probability of at least 0.75 sends the report to Escalate. Otherwise, high severity or low confidence sends it to Review; the remaining reports go to Monitor. With saving enabled, the script attaches the judgments to the PDF, uploads a decision note and moves the file.
A task for a person is optional and created only for Escalate, not Review. The file moves before that task is created, so this is not approval before action. The demo offers a triage signal, not a final security or legal judgment. Its cutoffs need testing on representative incidents before real use.
Key takeaways
- Ask separate questions about incident type, severity and escalation instead of relying on a prose summary to choose the next step.
- Keep routing rules visible in code. A model's confidence estimate is not proof that its judgment is right or permission to act.
- Local-file mode avoids Box access but still calls hosted Jev. It is not an offline test, and the README's named Markdown fixture is absent from the inspected repository.