Dhairya Desai

About

ML/NLP engineer with healthcare and industrial IoT experience: built an Optum pipeline that converted 2M+ physician notes into structured entities and linked them with claims/pharmacy data to create an actionable patient timeline. Deep hands-on expertise in production NER, entity resolution, and hybrid search (Elasticsearch + embeddings/FAISS), plus robust data engineering practices (Airflow, Spark, data contracts, auditability) and experimentation-to-production rollout via shadow mode and feature flags.

Experience

Education

Key Strengths

• Built NLP pipeline structuring 2M+ physician notes into facts linked to claims/pharmacy for a patient timeline
• Hybrid ML + rules NER with section awareness, abbreviation expansion, negation and temporality handling
• Strong evaluation/quality loop: clinician-labeled gold sets, inter-annotator agreement, per-entity precision/recall tracking
• Active learning workflow with selective abstention and reviewer feedback into retraining
• Production-grade entity resolution design: deterministic + probabilistic linkage with blocking, thresholds, and human review band
• End-to-end traceability via audit tables/data lineage for linkage decisions
• Handled complex industrial time-series/sensor data issues (clock drift, unit mismatch, tag renames, duplicates) using Spark stateful windowing and watermarks
• Search relevance improvements using hybrid sparse+dense retrieval (BM25/TF-IDF + embeddings/FAISS) with cross-encoder reranking
• Reliable data workflow engineering in Python: Airflow orchestration, config-over-code, data contracts, idempotency/restart safety
• Built and deployed production ML/NLP system for hospital readmission-risk prioritization
• Strong data quality engineering for messy clinical data (validation, record-level flagging, manual review paths)
• End-to-end workflow orchestration with Airflow (DAGs, retries, logging, monitoring, alerting)
• Reliability-focused AI development (automated tests for code + data, monitoring, drift detection, alerts)
• Pragmatic model/retrieval selection (classical ML vs transformers/LLMs; lexical vs dense vs hybrid retrieval)
• Effective collaboration with non-technical clinical stakeholders; translated model outputs into usable workflows
• Delivered measurable operational impact (reduced unnecessary readmissions; saved nurses hours weekly)