I built a Telugu-English code-mixing classifier for my final-year project, and it failed in a way that redirected my career. It scored 89% on my test set and collapsed the moment my grandmother typed a real message — because my training data was scraped from Twitter, and my grandmother does not tweet. That gap between benchmark accuracy and a real speaker was my introduction to the actual hard problem in AI: not the model, but everything the model quietly assumes about the world it will meet. A master's in AI is where I intend to learn to close that gap rigorously.
My foundation is deliberately mathematical, not just applied. Alongside the standard machine-learning coursework I took electives in linear algebra, probability, and optimization, because I had watched too many peers treat neural networks as spells rather than functions. I reimplemented backpropagation from scratch before I let myself use a framework, and I read the attention-mechanism paper three times over a month, deriving the shapes by hand, before I trusted myself to fine-tune a transformer. That discipline is why my classifier's failure taught me something instead of merely disappointing me.
I am drawn specifically to low-resource NLP — the languages, like most Indian ones, where the billion-token corpora simply do not exist and the field's default assumptions break. This is where research still has room, where the interesting questions are about transfer, augmentation, and linguistic structure rather than scale, and where the beneficiaries are the hundreds of millions of speakers the frontier models serve worst.
After the degree I intend to pursue research — ideally toward a PhD — on multilingual models that work for Indian languages as first-class citizens rather than afterthoughts. My grandmother's failed message is a small thing, but it is exactly the kind of failure that a well-trained researcher should refuse to accept as normal.
For two years I have deployed other people's research, and I have reached the honest limit of that role. As a machine-learning engineer at a computer-vision startup, I took a defect-detection model from a paper to a factory floor, where it worked until the lighting changed at dusk and our false-negative rate quietly tripled. Fixing it required understanding domain shift at a depth my job never taught me — I patched it with augmentation heuristics that worked without my knowing why, and that not-knowing is precisely why I am applying for a research-focused master's.
My engineering record is strong enough to build a researcher on. I own our model-serving pipeline, I cut inference latency 60% by rewriting our preprocessing and quantizing the model, and I built the active-learning loop that reduced our labeling costs by a third. But every one of those wins was engineering around models I could not fundamentally improve, and in our reading group I am always the one who can implement a paper and never the one who could have written it.
I want the coursework I have been faking through blog posts: the statistics of distribution shift, the theory behind the robustness tricks I apply blindly, and the supervised research experience of contributing something new rather than deploying something known. I am specifically interested in robustness and domain adaptation for vision — the exact category of problem that broke my factory model at dusk.
My goal is a research or applied-research role where the job is to make vision models fail less catastrophically in the messy real world, with a longer ambition of a PhD if the research clicks. I have spent two years being the person who patches models. I am applying to become the person who understands them.