Selected Work
Isolating individual customer conversations inside eight-hour, multi-speaker audio recorded on wearables across retail stores.
Finding where one customer interaction ends and the next begins, by classifying what happened rather than who was speaking.
Screening retinal images for diabetic retinopathy severity, with visual explanations a clinician can actually audit.
Reading sentiment from speech and text together, because tone routinely contradicts the words.
Approach
Problem
Write down the cost of being wrong before writing any code.
What decision does this model inform, and what does each kind of error actually cost? A false positive and a false negative are almost never equally expensive, and that asymmetry should drive the metric, the threshold and sometimes the entire approach.
Writing
Why gluing two feature vectors together limits what a multimodal model can represent, and what the outer product buys you.
They get mentioned in the same breath and do almost nothing alike. A short note on which one to reach for.
On ordinal labels, imbalanced classes, and why the metric you optimise is a modelling decision rather than a reporting one.
Experience
YoyoAI · Bengaluru
Extracting customer–salesperson interactions from 8-hour multi-speaker audio captured on wearable devices across retail stores.
Toolkit
Contact
Open to conversations about machine learning engineering roles, interesting problems, or anything on this site you'd like to dig into.
