Senior AI Engineer at HealthTap

Developed agentic clinical-intake workflows using LangGraph and Pydantic to gather symptoms and summarize chart context, utilizing function calling to surface structured visit briefs for physicians. Built HIPAA-aware retrieval and grounding pipelines using Amazon Kendra for semantic search and Amazon S3 for secure document storage, improving consistency through reranker optimization. Trained and evaluated medical text classifiers on Amazon SageMaker, implementing threshold calibration and safety review to manage model abstention on ambiguous triage cases. Implemented rigorous LLM evaluation harnesses using agentic LLM-as-a-judge patterns to detect hallucinations and tool-use errors, creating release gates within AWS CodePipeline. Partnered with clinicians and security teams to translate care-delivery requirements into measurable model objectives and audit-ready AWS deployment practices. Optimized conversational intake flows by leveraging LangSmith for tracing and observability, compari

AI Consultant, Manager at Accenture Applied Intelligence

Architected retrieval-augmented assistants using Azure OpenAI Service and Azure AI Search, combining prompt orchestration with custom guardrails to shorten analyst research workflows. Led data engineering for document ingestion and vector indexing using PySpark on Azure Databricks, improving recall and freshness across multimodal knowledge bases. Evaluated LLM applications using task-specific golden sets and rubric-based review, establishing measurable acceptance criteria before production deployment. Built generation services with FastAPI and Azure Kubernetes Service AKS, supporting secure deployment patterns for open-source models and client-controlled infrastructure. Introduced multi-agent orchestration in 2024 using LangGraph and AutoGen for planner-executor patterns and human-in-the-loop escalation, reducing brittle prompt chains in complex enterprise processes. Developed Semantic Kernel implementations for enterprise clients to map operational constraints and convert prototype fi

Senior NLP Engineer at Grammarly

Led transformer-based ranking models (BERT-family) for rewrite features, utilizing product telemetry and linguistic signals to increase suggestion acceptance rates across high-volume writing surfaces. Designed large-scale text cleaning and weak-labeling pipelines using Spark and Airflow, generating training corpora from anonymized edits while maintaining privacy. Trained multilingual quality-estimation models with PyTorch and XLM-R, improving intent-preserving rewrite detection for global users. Built offline evaluation frameworks in Python using statsmodels and scikit-learn to measure precision, acceptance lift, latency, and model calibration. Collaborated with product analysts to segment usage, diagnose failure modes, and link NLP model quality directly to user retention.