Sivasri Laveti

Designed and deployed scalable LLM-powered microservices using FastAPI/Django, supporting high-volume document processing and RAG-based workflows; Built RAG pipelines using GPT-4o mini and ChromaDB with optimized chunking, retrieval strategies, and prompt orchestration, reducing HR queries by 40% across 300+ users; Engineered ETL pipelines for PDFs, images, PPT, Word, and Excel, improving extraction accuracy by 30% through optimized parsing and retrieval logic.; Integrated Gemini-based rule-extraction modules as modular micro-services, reducing manual review workload.; Designed and maintained back-end components (PostgreSQL, Redis, Docker, Celery, ChromaDB), improving system scalability and API response times.; Implemented GenAI integrations across services, including prompt logic, RAG routing, and vector-store retrieval, improving pipeline consistency.

Built backend services orchestrating GPT, Gemini, and Llama workflows with RAG and structured outputs; Fine-tuned Phi and Gemma models using QLoRA on two separate 20K-sample datasets for agriculture and water domains, achieving ~25% accuracy improvement through domain-specific instruction tuning and data curation; Developed async multi-agent workflows using LangChain and AutoGen with intent-based query routing

Tailored a Shazam-style audio fingerprinting backend using spectrogram peak hashing for fast matching; Built ML pipelines for data ingestion, preprocessing, model execution, and evaluation.