Building ML systems that actually work, from 99.7% accuracy deepfake detection to OCR pipelines that read handwriting better than most humans. Currently breaking things at CDAC, fixing things at AARVAK.
Building an OCR + automated grading pipeline using TrOCR and LayoutLMv3 to read handwritten answer sheets , because nobody should have to decipher exam handwriting by hand in 2026.
Ran a 50+ member tech community: led a 30-person core tech team across ML, web, and DSA, shaped the curriculum, and turned "let's do a workshop" into actual workshops, hackathons, and speaker sessions that happened.
Built and shipped ML models for EdTech - NLP pipelines, data preprocessing, and HR salary prediction systems that moved real accuracy numbers on real datasets, not just notebooks.
A fully functional neural network built with NumPy only — no PyTorch, no TensorFlow. Implements forward pass, backpropagation, gradient descent, and custom activation functions.
View on GitHub →Custom CNN trained to detect AI-generated and manipulated audio. Uses mel-spectrogram analysis and attention mechanisms to distinguish authentic from synthetic voices at 99.7% accuracy.
View on GitHub →Built a carbon data validation engine for EU CBAM compliance, achieving 97.4% true positive rate and 0.912 F1 on anomaly detection across 50K synthetic industrial records at 14.3ms per submission.
View on GitHub →
