I built Faculty Finder, an individual semantic search project that organized 100+ faculty profiles into a searchable system. I developed hybrid semantic and keyword search and deployed an API on Render with an interface on Hugging Face Spaces.
For the co-authored FASP project, I developed Frequency-Aware Structured Perturbation for identity suppression while retaining visual utility. I evaluated it against LFW and CelebA baselines and compared it with PrIdentity and PRISM.
In my OTT Viewer Drop-Off Analysis project for IIT Guwahati, I analyzed more than 33,000 episodes and built churn-prediction models that achieved 85% accuracy. I also created Tableau dashboards and proposed personalized recommendations for retention.
I'm currently designing a post-hoc correction framework for medical speech-recognition errors, using anchored CTC re-verification and calibrated abstention to flag low-confidence corrections for human review. My work draws on data science, statistics, and machine learning, including time-series forecasting and quantitative trading analysis.
