Sidhant Manale

Architected production RAG system combining BM25 sparse retrieval and dense embeddings with query expansion, reranking and conversational memory achieving 40% improvement in retrieval relevance and 25% increase in user engagement. Engineered LLM prompt optimization pipeline leveraging Chain-of-Thought reasoning and Few-Shot learning with examples crafter with SME guidance, achieving 35% accuracy improvement in regulation document analysis and urban planning applications. Reduced response latency by 78% and memory footprint by 71% through connection pooling, efficient data structures, and pipeline refactoring. Built automated LLM-as-a-Judge evaluation framework using GPT-4.1, eliminating 80% of manual annotation overhead.

Engineered RESTful API services using FastAPI to automate web scraping of 10,000+ legal case files from Indian courts, improving legal data accessibility by 60%. Architected Multi-Agent AI system using LangGraph for automated contract analysis, identifying adversarial clauses with 85% precision. Implemented production-grade RAG architecture combining BM25 and dense vector search for legal Q&A chatbot, achieving 40% improvement in answer relevance.

Built end-to-end time series forecasting pipeline using XGBoost with feature engineering, achieving 98.8% prediction accuracy on 10,000+ row dataset. Developed hierarchical taxonomical classification datasets spanning 5 vertical domains with custom ontology structures.