Research
Research was how I got into technical work. I started with natural-language processing in college and later moved into biomedical data. It is now one part of my role, but the habit of testing assumptions has stayed with me.
Introduces a schema-less retrieval system that interprets dataset text, harmonizes terminology, and ranks GEO datasets against natural-language queries with evidence for each match. On a curated benchmark, it achieved more than 90% precision and its largest gains on complex, real-world searches.
Evaluates GPT-4 against traditional cell-type annotation methods across nine public single-cell RNA-seq datasets from diverse tissues. Literature-backed retrieval improved annotation quality and granularity, but ambiguous and less-characterized cell types remained difficult, showing that GPT-4 is not a universal solution.
Proposes a multi-task learning framework that jointly learns three biomedical named-entity recognition tasks. The shared model reduced inference time by a factor of three without reducing prediction accuracy.
Studies fiction and non-fiction classification for paragraph-length text using interpretable lexical, syntactic, and part-of-speech features. The feature-based model substantially improved over a simpler baseline, while BERT reached 97–98% accuracy and the interpretable model revealed measurable differences in lexical density, character diversity, and word order.
Selected projects from my BS–MS at IISER Bhopal.
Studied syntactic and lexical signals in paragraph-length text. The work led to the COLING publication listed above.
Investigated rules based on dependency relations for extracting cause-and-effect pairs from English sentences.
Used Markov chains and part-of-speech transitions to study differences between fiction and non-fiction. Read the thesis ↗
Studied sentiment and writing style in news, and built the backend for an application that categorized articles. View NewsChase ↗