Hello! 👋
M.Sc. Mathematics @ IIT Delhi · Machine Learning · NLP · GenAI
I’m an M.Sc. Mathematics student at IIT Delhi, building ML systems across computer vision, NLP, and retrieval-augmented generation. I care about the work behind the metric: strong baselines, data-leak checks, reproducible pipelines, and evaluation that reflects the real task — including the runs that make my systems look worse.
Cited statutory QA over India's 2023 criminal codes — 1,059 sections of BNS, BNSS, and BSA. Every cited section is checked against the retrieved text in plain Python, and deleting the LLM faithfulness checker plus its rewrite loop moved RAGAS faithfulness from 0.309 to 0.543 at Recall@5 0.750.
Deep metric-learning verifier for genuine versus forged signatures. A suspicious 0.999 baseline exposed two dataset leaks; after rebuilding the split and pairing, the 17 MB model holds ROC-AUC 0.986 on unseen writers and 0.877 cross-dataset on NFI.
Text-only risk screener for the 13 CCPA dark-pattern categories plus a benign class, over 6,373 deduplicated UI strings. A leakage audit found 64.8% of test rows echoing a training row, so the split is grouped: fine-tuned DistilBERT scores macro-F1 0.883 there, while a 4 MB TF-IDF pipeline generalises better out of domain at 0.752.
Indian tech-salary predictor built as a reproducible ML delivery project. It cleans 110K records and deploys tuned CatBoost with MAE around ₹1.60 LPA, tracked with MLflow, DVC, and Docker.
Indian Institute of Technology Delhi
University of Rajasthan
Stanford University / DeepLearning.AI · supervised and unsupervised learning, recommenders, and neural networks
DeepLearning.AI · CNNs, RNNs, sequence models, attention, and transformers
IBM · LLMs, retrieval-augmented generation, agentic AI, and vector databases