Campaign Trust and Safety Triage Copilot
Personal project, 2026
An LLM-assisted review workflow for crowdfunding campaign moderation. The model recommends, a human decides.
- Six-stage pipeline: code owns anything with a checkable answer, one LLM call handles narrative interpretation, people own every final action.
- Every cited policy clause is validated against two registries; a citation quoting text that is not in the submission is discarded.
- Hand-labelled evaluation set of 51 cases with a 90% agreement bar and zero tolerance for missed sanctions-list hits.