computer-use agents by day at skan.ai. triton kernels by night. i learn by building the thing and getting it wrong first.
site · x · linkedin · cv · shaurya.m207@gmail.com
forge-kernelsfused triton kernels for fine-tuning, one call: forge.patch(model). every number comes from a committed result file, including where it loses.
|
pactcontinuous thought without the recurrent unrolling. reasoning resynthesised a stage at a time; 4–9× faster training than coconut. |
noctis · live demoa fair value for tokenised us stocks in the 48 hours a week nothing quotes them, with an error bar you can buy cover against. |
flyloom · code soonthe whole adult fruit-fly brain, 138,639 neurons, on one cpu core. fast enough that a lesion screen is a coffee break. |
septa · downloadhold a key, talk, let go. a 0.6b model on your machine cleans the transcript before it pastes. no cloud. |
trading harness · private for nowpaper-only autonomous trading research. llm agents write the journal and can veto a trade. they can't place one. |
writing — how hard it was to find open source to contribute to · things i got wrong about building hobitat
iit delhi ee · paper: glaucoma classification with yolo + swin, ieee isbi 2024 · the 50s reel









