research & interests
At Hesiod, I've spent my compute on generative models for audio and interpretability methods for vision-language models. I like to use sparse autoencoders to "study" how neural networks organize information "internally". Besides, my interests extend to computational neuroscience, particularly studying disease progression in Alzheimer's and related conditions.
Studying the inherent effectiveness of deep learning. Notable interests include disentangled representations, causal learning, topological methods, neural machine translation (at least in the past). My honest working method, most days, is to retrain with a different seed and squint at what changed.
group theory, category theory, algebraic topology, mathematical physics (Einstein field equations), type theory, ...
Tons of multi-rank training on GPUs of different large transformer models (especially since 2023/ 2024). Apache Spark, Dask, and k8s for building (and maintaining!) ML infrastructure.
Created some multi-platform (Android, macOS, Unity Web Player) games with Unity3D. Notable interests in pathfinding algorithms (Dijkstra's, A*) and diffuse shading (how objects reflect light).
Started tinkering with Android in 2010 (for the better, or worse), when programming for Android was a wild-west. Over almost seven years, built a couple of "non-trivial" apps and contributed to open source projects on GitHub, now part of apps published to the Google Play Store (at least the last time I've bothered to check, ca. 2015).
Work Experience at
