research Multi-Expert Hybrid Retrieval for Movie Recommendation M.S. Thesis — Learned query routing over semantic, lexical, collaborative, and emotion-aware experts for prompt-conditioned recommendation. Indian Multilingual Jailbreaking Built Indic-JailbreakBench and proposed four jailbreak techniques achieving up to 35% higher ASR in Indic languages than English. course Signboard OCR & Transliteration Multilingual OCR pipeline using Faster R-CNN and an attention-based GRU encoder-decoder for Hindi-to-English transliteration. 92% detection, 89% transliteration accuracy. Cross-Lingual Emotion Understanding for Indic Languages Teacher-student distillation of RoBERTa-GoEmotions into IndicBERT across 5 Indic languages — ~3.3× improvement over zero-shot baselines. self Direct Preference Optimization (DPO) from Scratch End-to-end PyTorch implementation of the DPO algorithm to align LLM outputs with human preference data. RLHF Pipeline from Scratch Implemented a full RLHF pipeline using Vanilla Policy Gradient (VPG) to steer a GPT-2 SLM to generate positive tiny stories. NutritionRAG — Grounded Nutrition Chatbot Full RAG pipeline with a web app — hybrid BM25 + ANN retrieval over a textbook corpus stored in pgvector, with inline citations.
