data scientist II at skan.ai · bengaluru
computer-use agents by day. triton kernels by night. i learn by building the thing and getting it wrong first.
things i built to find out. each clip is about seven seconds.
fused triton kernels for fine-tuning, patched into a hugging face model with one call. every number comes from a committed result file, including the shapes where it loses.
continuous thought without the recurrent unrolling. the model resynthesises its reasoning a stage at a time through a learnable query codebook, so training stops waiting on itself.
a fair value for tokenised us stocks during the 48 hours a week when no exchange or oracle says anything. published with an error bar, and cover priced off it.
the whole adult fruit-fly brain as a leaky integrate-and-fire model, fast enough on one cpu core that a lesion screen is a coffee break. you address neurons by name, not by 18-digit ids.
hold a key, talk, let go. a 0.6b model on your own machine turns the raw transcript into something you'd actually send. fillers gone, names and rupees left alone. no cloud.
the part that survives live markets isn't the agent debate, it's the evaluation machinery. llm agents write the journal and can veto a trade. they can't place one.
all at skan.ai so far.
leading a team on an agent that drives desktop GUIs across 12+ bfsi workflows. set-of-mark prompting and a thinker–grounder split got atomic steps to 95%.
next-activity prediction at 81% top-1 over 200+ classes, and the serving work that made it shippable: 2.9× throughput at 45% lower gpu cost.
distilled bert, bart and pegasus into domain models 6× smaller at 98% of teacher f1.
notes on things i build, break, and think about.