Sidhant Manale

Redesigned RAG architecture with hybrid dense–sparse search (BM25 + sentence transformers), multi-stage query expansion, neural reranking, and conversation-aware memory with history embeddings.

Enhanced LLM prompt engineering pipeline using Chain-of-Thought, Few/Zero-shot learning, domain personas, and expert knowledge integration, boosting accuracy in legal and urban planning tasks by 35%.

Delivered 40% higher response relevance and 25% uplift in user engagement through improved system architecture.

Improved ML system performance through architectural and algorithmic optimizations:

Cut application latency by 78% via model quantization and caching.

Reduced memory usage by 71% with optimized data structures and pipeline redesign.

Streamlined codebase by 45% using modular design and dependency consolidation.

Developed an LLM-as-a-Judge evaluation pipeline, reducing manual annotation effort.

Developed REST APIs to scrape legal cases from various Indian courts, enhancing accessibility to legal information.

Designed a Multi-Agent system using LangGraph to identify one-sided and red-flag clauses in legal contracts.

Implemented advanced RAG techniques like hybrid search and metadata filtering for a legal QA chatbot.

Scraped and formatted laws from the USA, Dubai, and Singapore to curate a dataset for LLM fine-tuning.

Built a taxonomical dataset for five domains tailored to company-specific needs.

Designed a time series pipeline using XGBoost with∼10,000 rows achieving 98.8% accuracy.

Evaluated open-source text-to-speech models to enhance chatbot capabilities.