At Department of CSE, BIT Mesra, I engineered an adaptive windowing pipeline for fall data with severe class imbalance, achieving an 81.6% F1-score. I also benchmarked TS2Vec against CroSSL and raised model specificity to 87–95% under extreme label scarcity.
At Future Shift Labs, I built a PyTorch convolutional autoencoder for deepfake detection using reconstruction-error thresholding. It achieved 88.2% accuracy on real and fake images.
I built SATYA, an AI-powered truth and digital forensics platform, using LangGraph to coordinate adversarial agents and a verification shield to ground responses in real-time sources. I also developed MemoCare’s MRI and clinical-data diagnosis pipelines, deployed as a Flask app with PDF reporting and MongoDB authentication.
Through contributions to sktime and sktime-mcp, I implemented structured error boundaries for LLM hallucinations, multivariate broadcasting for BaseDetector, and capability:update support for transformers. My other projects include multimodal fashion retrieval and an NLP pipeline for exploring depression-related themes in Indian Reddit posts.
