MTech AI · IISc Bangalore
AI researcher working at the intersection of computer vision, model security, and responsible autonomous systems. Currently at Mercedes-Benz R&D India.
AI Research Scientist at MBRDI, working on autonomous driving perception, fairness evaluation, and LLM security.
Conducted the first power side-channel attack on SMAUG-T, the Korean post-quantum cryptography KEM standard, using ChipWhisperer-based trace acquisition and leakage analysis on embedded implementations.
Demonstrated key extraction with >99.9% success probability using fewer than 10 power traces — a result with direct implications for real-world deployment of KpqC-standard devices.
Fine-tuned Qwen3-0.6B with Unsloth and LoRA on CodeXGLUE tasks — defect detection, clone detection, summarization. Merged task-specific adapters into a unified multi-task model, then compressed via Wanda pruning, post-training quantization, and SVD-based FFN decomposition.
Hybrid retrieval pipeline combining BM25, FAISS-based dense vector search, and Reciprocal Rank Fusion (RRF). 300K 768-dim embeddings via Sentence Transformers. Best NDCG@10: 0.9356. Best MRR: 0.9167. Extending with TurboQuant-based vector compression.
Conversational Text-to-SQL assistant fine-tuned on Phi-3 Mini 4K using LoRA and Unsloth. 94.0% Exact Match Accuracy. 91.22% Average Token F1. SQLGlot-based validation and self-correction achieving 100% syntactically valid SQL generation.
Evaluated InternLM2-Chat on NovelQA (200K+ token contexts). Implemented and compared RAG, Multi-hop Event Graph Reasoning, and HyDE retrieval strategies. Improved accuracy from 45.45% to 54.54%. Optimized inference with LMDeploy; extended evaluation using Gemma-3 27B pseudo-gold answers.
Indian Institute of Science (IISc) · Under Prof. Venkatesh Babu
University of Engineering and Management
* IISC BANGALORE · AI RESEARCH SCIENTIST · MBRDI