Hi! I'm Dani, a researcher and engineer from the Canary Islands, currently living in New York City.
I work at the intersection of AI safety and mechanistic interpretability, with a particular interest in understanding latent representations in large language models. I also care deeply about the immediate societal risks that AI poses in areas such as mental health, democracy and critical thinking.
My work in the field has covered the geometry of belief states in transformer models and causal dependencies between sparse autoencoder features. Previously, as a founding ML researcher at Plastic Labs, I built systems for high-fidelity user representation in LLMs. I'm drawn to questions about what LLMs actually "know", how that knowledge is organized, and what that means for alignment, and enjoy writing short notebooks illustrating concepts from first principles.
Before focusing on AI safety research, I spent several years in product leadership and software engineering. For six years at BMAT Music Innovators, I worked first as a software engineer and later founding and leading the US team, launching key music rights products for streaming platforms, and managing relationships with SoundCloud, ByteDance/TikTok and Warner Music, among others. In my college years, I studied deep learning, signal processing and probability theory and their applications to language, speech and music. For my Bachelor's Thesis, I explored recurrent neural networks for speech detection during a research internship at Telefónica.
A through line connecting most of my work is the question of how technology mediates cognition, curiosity and creativity, a preoccupation that has driven projects ranging from AI-enhanced walking meditations and personalized digests to experimental instruments driven by gesture tracking and cellular automata, and low-tech and high-tech approaches to music discovery.