My work's all about the intersection of model behaviour, user safety/redteaming and taste. I care about measurable, actionable ways for video, image, audio and coding models to perform even better in a world that refuses to be simple.
I built multimodal and safety evals as well as policies for multimodal models and AI products at ByteDance's Seed Research Lab, after developing APAC's first AI safety evals for AI apps in both Western and Asian app stores as a regulator.
My AI safety work began in 2021 on trust and safety content moderation systems at TikTok, before consulting for Big Tech and early-stage startups on AI safety and policy at Plug and Play's incubator. I like to operate where research, product and GTM intersect.
Evals
My work is shaped by my journey from Big Tech lobbying and setting AI regulations, to working with early stage startups and shipping for global audiences in a frontier model lab.
Safety policies and evaluations for global model launches (e.g. Seedance), including a wide range of products such as chatbots, virtual assistants, and image/video generation apps. Focused on model personality, user harm, image/video quality, agentic capabilities and potential abuses of models.
Regulatory work and redteaming to build APAC's first consumer standards for AI apps, focused on clear expectations for user protection, transparency, and keeping pace with new agentic/AI research developments.
A daily benchmark for Claude models that evaluates answers to current-affairs questions with web search. It measures how well a model handles fast-moving facts, contested reporting, and limits of certainty.
View repositoryEarly AI safety work on trust and safety content moderation systems at TikTok, followed by policy and safety consulting for large technology companies and early-stage AI startups.
SG->SF
Bridging 2 homes that teach new ways to think about technology and taste.
Bookshelf
Eliot Higgins
A book that maps onto my SOCMINT hobby: following traces, finding unlikely sources of knowledge, and learning new hacks or networks. It's the same curiosity I bring to evals: keep asking where the evidence lives, who can see or measure it, and searching for unknown unknowns.
Contact