New Delhi · Developer & researcher
I build local AI tools and study how language models behave. My work connects browser systems, multilingual NLP, and careful evaluation: what works, where it fails, and how to check it.
Systems to try. Experiments to inspect.
Browser systems · Model evaluation
A local JEE solver in the browser, with a documented record of fine-tuning and model conversion.
7/12 → 8/12 on reused development questions; procedure scores fell. No adapter promoted.
Read case study — LlamaJEE.js →Reproducible experiments
How language, text format, and tokenizer configuration change the space a prompt needs.
1,120 saved encodings agree between Python and JavaScript.
Read case study — Comparative Tokenizer Study →Applied NLP · Error analysis
What happens when a harmful-speech classifier meets a different dataset and label policy.
Three datasets, three seeds; the model ranking changes with the dataset.
Read case study — mBERT vs MuRIL →