Hey, i'm Marmik

welcome to my web eden.

since the past few years, most of my time is spent working on problems in deep learning. currently, i work on interpretability research, which i like to think of as the biology of artificial neural nets. i'm very interested in understanding how deep learning algorithms gives rise to its own world of beautiful structure and elegant complexity and behaviors.

this summer, i'm working at groq on compound ai systems and fast model inference. concurrently, i'm also exploring how to find mechanistic signatures of unfaithful reasoning in reasoning models using Cross Layer Transcoders.

in the past, i worked on independent research on interpreting Mixture of Experts (MoEs) for expert specialization which got accepted at ICLR'25.

i love working with my friends a lot so i worked on densetex with 3 of my best friends which is a fast and accurate model for image-to-latex for math equations using CNNs and transformers during buildspace and also was a ml engineer at a startup working on ocr using vision transformers. at my uni, i worked on computer vision systems in battlebots for competitive robotics tournament for more than a year and learned a lot about optimizing algorithms in hardware constraints.

apart from work, i enjoy reading books of various genres (my bookshelf) and i've also been painting for a long time (my paintings). recently, i've been learning to play guitar as well.

i like meeting new people and if anything from above is interesting to you, feel free to reach out to me on twitter or at marmik [at] berkeley [dot] edu