Sai Malla

Owned the end-to-end design and development of production-grade LLM and AI/ML applications. Built RAG systems with document loaders, chunking, embeddings, retrievers, reranking, and vector stores (FAISS, Pinecone). Integrated GPT-4o, Llama 3, and Mistral with LangChain using prompt engineering, structured outputs, query rewriting, context management, and model selection. Developed stateful agent workflows with LangGraph for retrieval, reasoning, validation, tool execution, state management, retries, and human-in-the-loop flows. Implemented MCP and tool calling to connect agents with APIs/databases/functions and added validation and execution error handling. Built FastAPI services with PostgreSQL/Redis and authentication for scalable agent APIs, and delivered evaluation using LangSmith and MLflow (latency, relevance, faithfulness, tool-call success). Used A/B testing and offline evaluation to compare prompts, retrieval strategies, and model versions; reduced latency by 25% and improved

Worked on a production recommendation platform using Python, Pandas, NumPy, PySpark, SQL, and Databricks/Delta Lake. Built and tuned recommendation, ranking, classification, and similarity models using scikit-learn, Spark MLlib, XGBoost, and LightGBM, improving relevance by ~15% through iterative optimization. Created reusable feature engineering and training pipelines with PySpark, Airflow, Databricks Jobs, and MLflow, reducing manual training effort by ~30%. Developed real-time inference services with FastAPI and REST APIs, using MongoDB and Redis caching to reduce API latency by ~35%. Advanced NLP and deep learning components with PyTorch/TensorFlow and Hugging Face Transformers (BERT and sentence embeddings) to improve understanding of products and user intent. Added semantic recommendation and vector-based similarity search. Used SHAP/LIME for interpretability and A/B testing for experimentation. Implemented MLOps with MLflow, Docker, Kubernetes, and CI/CD for experiment tracking