Multimodal incident-triage scaffold · Concept
IncidentSight
A safety-first FastAPI baseline that extracts log evidence, inspects image metadata and keeps every remediation step human-approved.
Result — 36 tests over a typed, validated pipeline; every remediation step is gated behind human approval and nothing runs automatically.
Problem — incident tooling that "reads screenshots" tends to overpromise, blur observed facts with guesses, and quietly take actions no human approved.
What I built — a safety-first FastAPI baseline for multimodal triage that extracts errors, exceptions, timeouts and HTTP status signals from logs, and validates an optional screenshot by reporting only its format, dimensions and size.
Architecture — observed evidence is kept strictly separate from hypotheses. The response includes suggested read-only checks and an action plan where every step is marked as requiring human approval; the service executes no remediation.
Tech — strict typing, validation, security limits and an explicit future provider port. True screenshot interpretation stays a documented next step behind consent, privacy review and evaluation.
