Gurgaon, India
May 2019 – July 2019
My first industry internship. I walked into Sprinklr's office in Gurgaon knowing how to train models on clean datasets, and left understanding why production ML is a completely different game.
Making Sense of Social Media
Sprinklr processes an overwhelming stream of social media content. My job was to help make sense of it: build systems that could read millions of posts and tell you what people actually thought about products and brands.
I built sentiment analysis pipelines using LSTM architectures. The interesting part wasn't the model architecture, it was the data. Social media text is chaotic. People write in fragments, use sarcasm, mix languages, and communicate through emojis. A model trained on clean text datasets struggles with real posts.
I learned to think about NLP differently here. It's not about perfect accuracy on benchmarks. It's about extracting useful signal from noise, being robust to the mess of real human communication. The keyword extraction feature turned out to be more valuable than the sentiment scores themselves: knowing what people talked about mattered more than knowing if they were happy or upset.
Learning to Compress
The models worked, but they were too large and slow for production. This led me into model optimization: pruning weights that didn't contribute much, quantizing to smaller data types, finding ways to keep accuracy while shrinking the footprint.
This was my first encounter with the production constraint that would define much of my later work: a model that's accurate but impractical isn't useful. I learned to think about efficiency as a first-class concern, not an afterthought.
The internship was short, but it shaped how I think about ML. I went back to school with a different perspective. The gap between research and production became something I wanted to understand and bridge.