Malaysian field-builder at the intersection of AI safety, civic infrastructure, and education — building open, multilingual evaluation tooling so the people affected by fast-moving LLMs can decide better.
I do technical AI-safety work — evaluations, a little mechanistic interpretability, field-building — so that the people actually affected by fast-moving LLM development can decide better. My real audience isn't other researchers; it's consumers, families, government agencies, and NGOs who currently have to make calls about AI tools without the evidence to back them. That's the "meta-overview" I work from: build the underlying picture carefully, so many downstream decisions get easier — not to hoard expertise, but to give it away.
I work this way on purpose. I believe in being frugal and restrained with resources, and finding joy in that restraint — a barakah-over-maximisation stance. It's also why I believe in the meta-overview: a small, well-made piece of evidence infrastructure can serve far more people than another model or another app. BankBench-MY — the open, multilingual safety evaluation for Malaysian AI banking agents I'm building as a Sinar Project fellow — is just one case study in that logic. It shows how evaluation can feed three things at once: policy (what BNM's RMiT framework should actually require), safe products for society (what banks should ship), and public education (what my mother, using MAE or GrabPay, deserves to understand). The method matters more than the domain.
I'm slowly teaching myself six languages at once — Arabic, Latin, Hindi, Korean, Hebrew, and German — mostly through films (Hindi via Bollywood is a personal favourite). On the side I draw Japanese-style line illustration (Noritake and Henn Kim are my muses) and am developing a healing-focused illustrated children's book series. I also tutor a handful of secondary students, which keeps me honest about what "education" actually feels like from the other side of the desk.
Shafira Noh is a Sinar Project Civic Tech Fellow who believes the best AI-safety work is done lightly — frugally, with restraint, and a real joy in using few resources to help many. She builds open, multilingual evaluation tooling, starting with BankBench-MY (a safety benchmark for Malaysian AI banking agents), so that consumers, families, government agencies, and NGOs can decide better about fast-moving LLM technology. Her approach is hands-on and quietly decorated: she was part of the winning team at the BlueDot–Apart Hackathon (AI Safety Public Education track) and has workshopped AI safety with 300 teachers at DutaGuru. Off the clock she's teaching herself six languages at once and drawing Japanese-style line illustration — because, for her, curiosity and care are the same muscle. She works at a restrained scale so the evidence can serve many, not few.