About
I've tried many mediums of creativity in my years. my father wanted to take art classes when he was younger but he couldnt afford it, so he really made sure i had no excuses.
i tried drawing and painting, playing the guitar, keyboard, translating comic books from english to persian, photography etc. i enjoyed doing all of these, however i was also very interested in exploration and soon discovered most things i were doing had been tried before. there are only so many ways you could make a chord progression sound good, or mix and place colors on canvas.
somewhere along the line i learned about transistors and that led me down this computing rabbithole. i decided that language and math were the frontier and it was the best use of my time is to pursue exploration in that.
this was my state of mind when LLMs happened. i remember some horror at the realization that the world had again changed violently in my formative youth years and that i'd have to adjust. more than that tho was the thrill of exploration. Here is something right at the edge of the uncharted, closely related to math, language and religion which all are important foundations in humans.
i decided that i need to learn all there is about machine learning and these stupidly complex systems. i've been having a blast tinkering ever since.
i now view these systems as mirrors. they show us patterns in how we think, communicate, and create value. every prompt is a negotiation, every output a reflection of collective human knowledge filtered through probability distributions.
i'm particularly interested in the gaps; the places where these systems break down or surprise us. that's where the real learning happens. when an AI hallucinates, it reveals something about information processing. when it solves a problem elegantly, it shows us new ways to think about intelligence itself.
my current focus is on the frontier between AI and physical systems. analog-gradients is the embodiment of this - an autonomous chip design exploration platform that closes the loop from PyTorch to silicon-grade timing, area, and power metrics in about 11 minutes. the first feedback loop between ML model design and physical silicon.
we live in a world where most people can now leverage the cumulative knowledge of humanity to create. that changes everything. the tools aren't the bottleneck anymore - vision is. i want to build the workshops that let small teams do things that used to require institutions.
i operate through calibur labs. the long-term vision is compute sovereignty - chip design should be accessible enough that intelligence infrastructure is distributed rather than centralized.
this site is where i document the build in real time. some posts are technical, some philosophical, most are somewhere in between.
reach out if you want to build something together or just talk about the weird future we're creating.
-M