Following an AI system across the whole stack
A map of the territory I want to explore—from distributed systems and reliability down to kernels, memory, and hardware.
Read noteData scientist · systems thinker · perpetual student
I’m Soham Joshi, a Data Scientist at Meesho and an IIT Bombay graduate. I’m interested in understanding AI systems across the entire stack—and in turning technical insight into useful, dependable products.
This is not a finished-person portfolio. It is a growing library of essays, learning notes, research ideas, and beliefs that may change when they meet better evidence.
Understand enough to build. Let reality expose the gaps. Return with better questions.
The library
Essays, research directions, and notes—newest first.
On self-teaching, following questions into uncomfortable depth, and learning to balance understanding with action.
Read essayA map of the territory I want to explore—from distributed systems and reliability down to kernels, memory, and hardware.
Read noteWhat becomes possible when capable models run privately and efficiently on ordinary devices?
Read noteCurrent threads
These themes will become essays, experiments, or discarded ideas.
Test-time training, continual learning, RDL, and systems that do not remain frozen after training.
Distributed systems, infrastructure, reliability, observability, and the journey from prototype to product.
CPUs, GPUs, memory hierarchies, kernels, compilers, and the architecture of local intelligence.
How insight becomes a product, how demand is discovered, and why distribution can matter as much as invention.
Chess, UFC, geopolitics, history, neuroscience, and psychology as windows into action under pressure.
Fitness, travel, discipline, and balancing a life of abstractions with the physical and human world.
About this place
I want this website to preserve the development of an idea, not only its polished conclusion. Some thoughts will become essays. Others will remain unfinished research notes. Both belong here.
More about me and the library