Larry H.

Led the development of a production-grade RAG platform facilitating natural language querying of millions of insurance documents. Automated ingestion, semantic retrieval, and LLM-based answer generation, reducing document analysis time significantly. Designed an end-to-end pipeline encompassing document ingestion, parsing, chunking, embedding, vector indexing, retrieval, and LLM inference. Built scalable data pipelines using Python and Apache Airflow for processing large volumes of unstructured and structured data. Engineered embedding and indexing workflows utilizing vector databases (FAISS/Pinecone/OpenSearch) for semantic search over 1M+ document vectors. Achieved query latency of approximately 38ms by optimizing retrieval pipelines. Implemented hybrid retrieval strategies and developed LLM-powered inference pipelines with OpenAI/Azure OpenAI. Constructed backend services and REST APIs (FastAPI, microservices) for real-time querying and summarization. Deployed cloud-native infrastructure across AWS and Azure leveraging Docker and Kubernetes (EKS/AKS). Established MLOps workflows, including CI/CD pipelines, model versioning (MLflow), and monitoring. Integrated observability and monitoring systems (Prometheus, Grafana) to ensure reliability and performance. Collaborated with frontend teams to support a chat-based UI for analyst interaction with datasets. Led a team of engineers, managing system architecture and delivery across multiple functions.

Constructed distributed ML infrastructure enabling real-time event detection from global data streams. Designed streaming pipelines that managed billions of daily inputs, including text, image, and sensor data. Enhanced system throughput through effective pipeline optimization. Automated model deployment, reducing time from weeks to under one week via CI/CD practices. Developed feature pipelines supporting both real-time and batch ML workflows. Established model serving systems for low-latency inference at scale. Implemented monitoring and alerting systems (Prometheus, Grafana, ELK), leading to a notable reduction in incident detection time.

Developed ML models, including logistic regression and random forest, aimed at predicting telecom customer churn. Designed data processing pipelines for large-scale CRM and billing datasets. Generated risk scores for millions of users to assist in retention strategies. Enhanced model performance through rigorous feature engineering and evaluation. Deployed batch scoring systems tailored for enterprise ML workloads.