Sai B.

● Built Verizon OpenAgentPlatform, an enterprise AI platform for teams to compose LLM agents, MCP tools, and RAG (Retrieval-Augmented Generation) collections into reusable chat and API experiences, enabling 1700+ users to launch AI workflows without bespoke engineering. ● Implemented LangGraph multi-agent architecture with thread-ID conversation management and crash-safe summarization, increasing successful long-running conversations by 60%. Integrated MCP (Model Context Protocol) tools and RAG collections (LangConnect) with pgvector DB to power semantic search and retrieval over enterprise content with standardized tool schemas and guardrails. ● Prototyped and evaluated LLM flows using Hugging Face Transformers (Llama, Mistral) and hosted LLM APIs, comparing latency, quality, and cost before baking them into the platform’s agent presets. ● Engineered enterprise governance with RBAC, subscription-based access, sensitivity-level visibility, end-to-end audit trails for tools/RAG/agents, satisfying internal security and compliance requirements for AI usage. ● Added configurable agent controls (LLM, temperature, tokens, system prompts) with AI-assisted risk scoring and human-in-the-loop approvals reducing risky configurations by ~35%. Integrated Langfuse LLM tracing and observability dashboards (DAU, new users, completion rate, latency, engagement) to identify failing prompts, optimize routing, guide product decisions to improve agent completion rate by 15%. ● Designed a Promote to Production workflow that triggers Jenkins pipelines to migrate entities from UAT to PROD at the database level, enforcing environment isolation while enabling fast iteration; integrated with GitLab CI/CD, Docker images, and internal deployment standards. ● Developed and maintained the platform’s backend and UI using Python, FastAPI, Postgres, Node.js, React, Swagger/OpenAPI, Docusaurus, and internal ADOM tooling, ensuring consistent APIs, documentation, and developer experience.

● Partnered with data science and product teams to build and productionize machine learning models for demand forecasting and replenishment, improving forecast accuracy by 4% and contributing to a 3% reduction in stockouts and overstock across targeted categories. ● Designed and implemented data and feature pipelines on Spark and Databricks, using PySpark, SQL and Delta Lake to transform clickstream, transaction and catalog data into reusable feature tables, cutting pipeline failures by around 30% and reducing data lag for training and batch inference from a day to a few hours. ● Developed models in PyTorch, scikit-learn with NumPy and Pandas for time series forecasting and classification use cases, and tracked experiments, parameters and model versions with MLflow, shortening iteration time for new model variants by roughly 25%. ● Deployed models as containerized FastAPI services on AWS using Docker, Kubernetes and GitLab CI/CD, consistently meeting p95 latency under 250 milliseconds for online inference while supporting several hundred requests per second during peak traffic. ● Co-developed a generative AI assistant for supply chain and planning teams using retrieval augmented generation over internal documentation backed by a vector DB and FastAPI backend, reducing manual dashboard and report lookup time for pilot users.

● Contributed to machine learning solutions for large scale e-commerce use cases such as search ranking, product recommendation and delivery time prediction, using Python, Pandas, scikit-learn and gradient boosted tree models like XGBoost and LightGBM to drive click through and conversion lifts in the range of 2-3% on targeted pages. ● Built and maintained data preprocessing and feature engineering pipelines with Spark, PySpark and SQL to process logs and catalog data at tens of millions of events per day, reducing data quality issues and failed training runs by around 30%. ● Helped deploy trained models as Docker based services on internal orchestration platforms, integrating them with existing microservices, logging and metrics to keep low latency and error rates under 1% in production. ● Documented data contracts, model workflows, operational runbooks and participated in internal knowledge sharing sessions, helping cut onboarding time for new engineers and improving consistency of practices across projects.