Safety research engineering for AI harms that affect people.
My work focuses on where AI systems can enable abuse, distort evidence, or make high-stakes decisions harder to trust. I build tools, evals, and prototypes to mitigate those risks.
Experiments
- Best-of-N StrongREJECT eval An LLM refusal-robustness eval measuring how Best-of-N adversarial prompt search scales against gpt-4.1-nano, with query accounting, confidence intervals, and redacted public artifacts. LLM refusal · Evals
- Text-only T2I jailbreak eval A text-to-image safety-filter eval separating filter bypass from moderation-confirmed harmful success under SneakyPrompt/SurrogatePrompt-style text attacks. T2I filters · Evals
Tools
- What is this image? A browser-only image-authenticity inspector. It reads C2PA provenance to show who made a picture, with what tool, and whether AI was involved. Provenance · C2PA
- Media server A self-hosted Jellyfin media server I run on a VPS, with the Radarr, Sonarr and Prowlarr stack behind Traefik. Ansible bootstraps the server and all services run in their own Docker container. Self-hosted · Ansible