shaurya madukuri — data scientist, skan.ai
right now that's computer-use agents at skan.ai. on the side i write triton kernels to make training cheaper and cheaper. i learn by building the thing and getting it wrong first.
now — leading a small team on desktop computer-use agents for BFSI.
nights — triton kernels for memory-efficient training; contributing to unsloth.
reading — how model weights leak their training data.
i'm shaurya. electrical engineering from IIT delhi, now a data scientist at skan.ai. the title says "applied AI"; most days i'm just trying to get models to reason a little better and run a lot cheaper.
lately that's a computer-use agent — the kind that drives a real desktop GUI, screenshot by screenshot, for finance workflows. before that i did process discovery: next-activity prediction, fine-tuning small LLMs, and a fair amount of inference work (quantization, speculative decoding, KV-cache tricks) to make it cheap enough to actually ship.
the part i'd do for free is the low-level stuff: fused kernels, where memory actually goes during training, whether a reasoning trace has to be sequential at all. i also mess with lattice cryptography, which is a different way of caring about the same thing — making an idea small and exact.
leading a team on a desktop-GUI agent across 12+ BFSI workflows — 95% atomic-step success.
next-activity prediction, multi-model RAG, and inference work — 2.9× throughput at 45% lower GPU cost.
teacher–student distillation of BERT / BART / pegasus — 6× smaller at 98% of teacher F1.
latent reasoning without the recurrent unrolling — stage-wise resynthesis via a learnable query codebook. 4–9× faster training than coconut, ~5× memory reduction, scaled to llama-3b/8b.
triton kernels for memory-efficient training. fused linear + cross-entropy cuts peak activation memory 62% at 128k vocab; fused rmsnorm / swiglu / rope run 1.4× over the pytorch baseline.
a social network for slowly getting good at hobbies, in public. i got the shape wrong — a consistency-shaped product for an inconsistency-shaped behaviour. the postmortem →
also a paper: glaucoma abnormality classification using YOLO and SWIN transformer, IEEE ISBI 2024.