A huge amount of everyday text in India — chats, comments, captions — mixes Hindi and English in the same sentence, often in the same word. Most off-the-shelf NLP tooling is trained on clean, single-language text and quietly breaks on this.
This isn’t a niche problem. It’s the default way a large population actually writes. Treating it as an edge case means building tools that don’t work for the people who’d use them most.
We think this is one of the more practical, underexplored directions in applied NLP right now: not chasing bigger models, but building tokenization, sentiment, and intent-classification approaches that take code-mixed text seriously as its own thing, rather than a broken version of English.
It’s slower, less glamorous work than fine-tuning the latest model on a benchmark. It’s also the kind of work that actually changes whether a tool is usable for the person it was supposedly built for.