What it is: Bhojpuri sentiment resources built from AI-model disagreement.
Why I built it: Bhojpuri has tens of millions of speakers and almost no labelled data.
What I did: I built the pipeline, collected the text, ran the open-model committee on free GPUs, and released a model and a dataset.
My role: Lead. Prof. Deepak Kumar Rakesh (NIT Bhopal) mentored the work. Dr. Amit Kumar joined through Prof. Rakesh and reviewed an early draft, then left the project voluntarily.
What it is: A bare-metal Rust kernel that checks a file's integrity every time it is read, not just once at boot.
Why I built it: A file can be changed after it has been checked.
What I did: I built on Philipp Oppermann's blog_os tutorial. I learned Rust in 2024 and started the kernel offline in 2025. I added per-read checking, then per-block Merkle checking.
Design guidance: Prof. Subodh Kumar (IIT Delhi) guided the design by email over about two months. He pointed me to profiling with hardware performance counters, a Merkle tree, and checking per block with lazy updates. I wrote all the code.
Patent: An Indian patent application covers the per-read checking method.
What it is: A study, in progress, with a PhD collaborator. Across 9 tokenizers on FLORES-200, Bhojpuri costs about 1.2× and Maithili about 1.3× as many tokens per word as Hindi under Indic-trained tokenizers.
Why I built it: Bhojpuri and Maithili cost more tokens per word than Hindi, which means more compute and money for the same sentence.
What I did: Measured how many tokens each word costs in Bhojpuri, Maithili, Hindi, Marathi, Nepali and English, across 9 tokenizers, on the 1,012 parallel sentences of FLORES-200. Checked that the languages have the same characters per word, so the gap isn't caused by longer words.
My role: I built the measurement scripts and ran the analysis.