I Tried to Teach a Machine to Read Hindi Feelings

The Hindi Sentiment Analysis project, pushed in November 2024, was personal before it was technical. Mental health conversations in India often happen in Hindi or Hinglish, yet so many NLP tools pretend the world speaks only English. I wanted to build something that could classify text as positive, negative, or neutral with respect for that reality. The Python pipeline combined preprocessing, model experimentation, and evaluation metrics I actually understood — not just accuracy numbers I could quote in interviews. There were frustrating weeks where the model confused sarcasm with sentiment, and victories when a batch of test phrases finally classified correctly. I kept pushing updates into 2025 because the problem felt unfinished in a meaningful way, not because a syllabus demanded it. This repo marked my pivot from apps that work to systems that interpret language. It also prepared me for heavier NLP work later, like legal document analysis. If RefactorAI is about code language, this project was about human language — and that distinction still guides how I choose problems worth solving. Building for your own context beats copying Kaggle notebooks every time. I read papers, tried simpler baselines, and learned to report failure modes honestly. That habit showed up later in legal NLP and in RefactorAI evaluation scripts — same engineer, harder problems. If I could tell past-me one thing, it would be this: start with the language your users actually speak, then chase the metrics.