Satvik G S

Engineered a scalable Python-based data ingestion pipeline, scraping over 50,000 pages from diverse sources to support a Retrieval-Augmented Generation (RAG) system.

Designed a custom chunking algorithm, extracting structured data from over 12,000 documents (PDFs, PPTs, DOCs) including text, tables, and images.

Optimized LLM context retrieval and MilvusDB ingestion using Delta run and Re-config optimizations to enhance system performance.

Fine-tuned hybrid, keyword, and semantic search models with re-ranking techniques, boosting search relevance scores by 25%.

Benchmarked and optimized open-source LLMs (LLaMA 3.1, Mistral 7B, Mixtral 8x7B) through LoRA, QLoRA, and PEFT fine-tuning techniques.

Evaluated proprietary LLMs (OpenAI, Claude, Gemini-1.5 Pro/Flash) to inform cost-effective deployment strategies.

Automated information retrieval, reducing processing time from 1 hour to under 30 seconds, achieving an 85.7% efficiency improvement.

Evaluated Neo4j database integration into existing architecture, extracting relevant table and column descriptions from more than 50 tables.

Engineered core logic for a ReACt Agent using the LangGraph framework, enabling dynamic tool selection based on user queries.

Jointly led the integration of Google's Gemini LLM and few-shot learning, elevating SQL query generation accuracy from 70% to 90%.

Revamped the backend codebase, decreasing overall response time from 27 seconds to 20 seconds, achieving a 25.9% improvement.

Received an Amazon voucher award for exceptional contributions to the project.