SAI PREETHAM

Built and shipped production RAG pipelines for internal financial knowledge workflows across 100K+ enterprise documents, improving retrieval precision by 30% and relevance/recall by 25%+ through experimentation and failed-retrieval analysis. Architected and deployed production multi-agent workflows on LangGraph using state schemas, conditional routing, checkpointed execution, and a human-in-the-loop approval process coordinating 7+ specialized agents. Extended agent capabilities via reusable MCP tool servers that connected to internal services through schema-validated interfaces with authentication and permission controls. Benchmarked cross-cloud LLM deployment trade-offs using Oracle OCI GenAI expertise and AWS Bedrock, then deployed production GenAI services with Anthropic Claude. Implemented automated LLM/agent evaluation frameworks (1,000+ test queries) with groundedness/hallucination/retrieval and tool-call accuracy metrics, wiring pass/fail thresholds into regression gates to cut

Trained and benchmarked 15+ NLP/ML configurations using Python, PyTorch, TensorFlow, and scikit-learn, achieving 0.86 F1 via feature refinement and robust evaluation (cross-validation, threshold analysis, and segment-level error analysis). Managed 50+ training runs using MLflow for reproducible experimentation, tracking hyperparameters, metrics, and artifacts. Built reusable data preparation/feature engineering pipelines for 1M+ records (Pandas, NumPy, SQL), reducing training dataset preparation time by 22%. Shipped Dockerized Python REST inference services with payload validation, model versioning, structured logging, exception handling, and health checks. Implemented monitoring/diagnostics to detect drift by comparing production feature distributions and predictions against training baselines, triggering timely retraining. Automated ML validation and regression testing with 6+ checks (data quality, feature consistency, inference outputs, performance thresholds), reducing manual valid

Prepared and transformed 500K+ structured records using Python, Pandas, NumPy, and SQL to build preprocessing workflows (missing-value handling, outlier treatment, categorical encoding, scaling, and feature transformations). Improved classification performance by 14% using feature selection, hyperparameter optimization, stratified cross-validation, and threshold tuning with metrics such as precision/recall, ROC-AUC, and confusion-matrix analysis to address underperforming segments. Exposed trained models via REST APIs with payload validation, integration testing, model versioning, structured logging, and deployment verification, establishing a repeatable path from experimentation to downstream consumption. Built transformation-consistent preprocessing pipelines to reduce training-serving discrepancies, and worked with software engineers to troubleshoot request schema and environment/preprocessing dependency issues during release cycles.