LLM Evaluation & Guardrails Standard
Defined how LLM-powered features are evaluated, red-teamed and monitored after release.
- ORGANISATION
- Earthstorm Media
- ROLE
- Engineer → Governance
- DATE
- May 2024
SafetyAssuranceOWASP Top 10 for LLMsNIST AI RMF
Overview
This is where my engineering and governance work met. I wrote a standard covering evaluation datasets, prompt-injection testing, output filtering and production monitoring, then built the first reference implementation myself.
Outcomes
- Standard adopted for all customer-facing LLM features
- Red-team checklist mapped to the OWASP LLM Top 10
- Monitoring dashboard tracking refusal and hallucination rates