Jaykumar P.
Generative Ai Engineer @HCA Healthcare
Signup · Get unlimited contacts
WORK HISTORY
Generative Ai Engineer @HCA Healthcare
US
Built an Agentic RAG clinical-documentation assistant using LangGraph with Azure OpenAI (GPT-4o) andAWS Bedrock (Claude 3.5 Sonnet) as fallback; orchestrated retrieval, tool-calling, and reflection nodes with human-in-the-loop physician review, cutting chart-summary turnaround from approx 12 min to approx 3 minand reclaiming approx 2,400 physician hours/month across 14 facilities.• Designed a HIPAA-aligned RAG knowledge base over 8M+ clinical notes and policy documents usingLlamaIndex for hierarchical chunking, open-source BGE-M3 embeddings (self-hosted on GPU) for PHIisolation, and Pinecone namespaces per facility; achieved recall@10 of 0.91 and lifted answer faithfulness from 78% to 93%.• Optimized inference latency and throughput for self-hosted clinical models (including the fine-tuned Llama-3) by migrating serving infrastructure to vLLM; leveraged PagedAttention and continuous batching to increase token generation throughput by 3.5x, slash p90 time-to-first-token (TTFT) by 60%, and reduce monthly compute costs by approx $22K.• Engineered a document-parsing pipeline for scanned lab reports and insurance forms using Tesseract OCR with layout-aware post-processing and LLM-based field extraction, automating ingestion of approx 22K documents/day at 96.4% field accuracy and eliminating approx 5 FTEs of manual data entry.• Fine-tuned Llama-3-8B with LoRA (QLoRA, 4-bit) on 180K de-identified discharge summaries for clinical note generation, achieving 94% summarization accuracy and ROUGE-L of 0.62, within 2 points of GPT-4o at approx1/20th the inference cost for high-volume batch workloads.• Implemented production guardrails (NeMo Guardrails + custom PHI/PII redaction and hallucinationclassifiers) and an MCP server exposing Epic EHR tools to agents under RBAC, blocking 100% of test-setPHI leakage attempts and cutting unsafe completions from 3.1% to 0.2%.
EDUCATION
University of South Dakota
Master's degree, Computer Science
ABOUT JAYKUMAR P.
I\'m an open-source contributor(geoai, vLLM, huggingface, transformer repos) and a senior Machine Learning Engineer with 5+ years of experience in a startup environment-Now focused on GenAI: specializing in designing and developing Python applications with LLMs and multimodal models, creating architectures aligned with generative AI behavior, implementing RAG pipelines, and integrating AI workflows with existing enterprise systems-Theoretical and practical implementation experience of machine learning techniques and algorithms: XGBoost, Graph Embedding, Linear and Nonlinear Optimization, Clustering, Regression, feature engineering- in-depth knowledge of Google Cloud Platform services, particularly Vertex AI, BigQuery, and Dataproc- Expertise with CI/CD and knowledge of distributed computing and big data technologies like Spark, Kubeflow, Airflow, and SQL- Extensive expertise in Python and machine learning libraries (e.g, TensorFlow, PyTorch, scikit-learn)-Advanced knowledge of software engineering principles, including software design, source control management, build processes, code reviews, and testing methods-Strong Python, SQL, Pyspark, and ElasticSearch skills. Strong analytical and data modeling skills. Advanced understanding of data pipelines and ETL processes-Strong knowledge of System Development Life Cycle (SDLC)-Demonstrated experience in organizing work and managing multiple potentially conflicting priorities while collaborating with stakeholders across various functions-Strong interpersonal, presentation, and communication skills and agile development experience.If you’d like to learn more about how my services can help your company, please reach out via Email (j••••••••@gmail.com)
This profile is compiled from publicly available professional sources. Unifers is not affiliated with or endorsed by LinkedIn. Request removal of this profile.