Pooja Dokuri

About

Built and deployed a production LLM + vector search clinical decision support system at UnitedHealth Group, retrieving medical evidence and patient context in real time for prior authorization and risk scoring. Strong in end-to-end RAG architecture (Hugging Face embeddings, Pinecone/FAISS, SageMaker, Redis) plus orchestration (Airflow/Kubeflow) and rigorous evaluation/monitoring, with demonstrated ability to align solutions with clinical operations stakeholders.

Experience

Education

Key Strengths

• Built and deployed production LLM + vector search pipeline for clinical decision support
• Designed high-throughput real-time feature/data pipelines (Kafka + Spark)
• Handled noisy/inconsistent EHR data via normalization, deduping, and clinical entity mapping
• Implemented hybrid retrieval (BM25 + embeddings) to improve relevance with incomplete text
• Production optimization for low-latency retrieval at million-scale (caching with Redis)
• Hands-on orchestration across Airflow, Kubeflow Pipelines, and SageMaker Pipelines
• Reliability-focused evaluation approach (offline tests, A/B testing, monitoring/drift tracking)
• Effective collaboration with non-technical clinical stakeholders; translated needs into ML requirements