Ajay Agarwal

Ajay is a Research Engineer at Citadel AI, conducting frontier model evaluations and field strategy for Japan's AISI and Cabinet Office. He is also part of Shiba AI Lab, a frontier AGI research lab in Tokyo, where he works primarily on model metacognition and technical AI governance.

​

Outside Japan, he collaborates independently with a team in the Netherlands on LLM steganalysis and with safety researchers in India. He is also a mentor at Algoverse, and volunteers with PauseAI and in the field-building run by AI Safety Tokyo and Shiba AI Tokyo.

​

Much of his work is aimed at supporting middle powers in scaling their governance frameworks to match the growing adoption of frontier models

​

Kyle Gabriel Reynoso

Kyle is a Research Manager for SASH’s AI Safety Fellowship, a three-month full-time fellowship in Singapore focused on technical AI safety and governance.

Before this, he was Learning Director at WhiteBox Research, where he led technical sessions and mentored fellows for two cohorts of the SEA-focused AI safety research fellowship. He also led program design for Condor Camp, Southeast Asia’s first AI safety camp in 2024.

Recently, he was a fellow at Pivotal Research during the Summer '25 cohort, investigating interpretability approaches to find concept boundaries in LLMs. He also worked on probe subversion evals under the UK AISI bounty program, and many-shot jailbreak interpretability at SPAR in 2024.

​

Tania Sadhani

Tania spent the last 5 years working in the emerging tech ecosystem and believes that a multidisciplinary approach is needed to responsibly develop and adopt high-risk emerging technologies.

She currently works at Generality Labs on Inspect Evals, improving open-source LLM benchmarks and supporting AI evals research.

​

She studied computer science and data analytics at the Australian National University, where she was a college STEM mentor. She has worked on researching Critical Learning Periods in Deep Learning Models, modelling AI Security risk and building MLSecOps processes, and shaping Australia’s Quantum Industry. She also completed Tara with a capstone on measuring LLM visual reasoning.

​

Oscar Yasunaga

Oscar is a physics and math student at the University of Toronto, where he has worked under Professors Zhijing Jin, Chandra Gummaluru, and Dennis Fernandes. His work spans AI safety and education. On safety, he is interested in mechanistic interpretability and sycophancy. On education, he builds interactive visualizers to make ML concepts easier to understand.

​

He also serves as a teaching assistant for the Toronto AI Safety Initiative (TAISI) Summer Intensive, where he has taught emergent misalignment, steering, evaluation science, and alignment faking.

​

The AI risk that concerns him most is the double-edged nature of open weight models, which counterbalance the concentration of power but can also be an attack vector.

​

Janhavi Khindkar

Janhavi is an Applied AI Researcher at IIIT Hyderabad, where she leads LLM training, optimization, and deployment for low-resource Indic languages as part of Bhashini, India's national multilingual AI program.

​

She founded ValueShift Research, an independent group spanning mechanistic interpretability, evaluation, and control, where her work decomposes LLM refusal behavior into SAE-verified circuit directions. She is also a PRISM AI Safety Fellow.

​

She has mentored at TalentSprint (IISc) and judged multiple Apart Research sprints, including the Global South AI Safety Hackathon, where she spoke on the multilingual blind spot in AI safety. She previously worked on agentic sabotage and eval awareness at Algoverse.

​

Isac Sahlberg

Isac is a computational physicist by training. During his PhD studies, his research in condensed matter physics has focused mostly on the robustness of topological phases in two-dimensional lattices with random geometries.

​

He has worked as a teaching assistant in mathematical methods and quantum mechanics courses, and he has facilitated the technical track of a BlueDot-based AI Safety Fundamentals course organized by the Finnish Center for Safe AI (Tutke).

​

Isac has helped organize an AI safety reading group in Helsinki over the past year. He was previously a participant in the ARENA-based alignment programs FAEB organized by Tutke, as well as CAMBRIA organised by the Cambridge Boston Alignment Initiative.

​