Omid Jafari

Projects

Selected systems and research projects.

My work spans clinical NLP, applied AI, healthcare data systems, research software, and full-stack development.

Many of these projects were collaborative research efforts. Each card describes the part I personally designed, developed, or substantially contributed to.

NLPMed Portal interface for reviewing longitudinal clinical notes and running clinical NLP workflows.

NLPMed Clinical NLP Platform

An end-to-end clinical NLP platform combining a Django portal with a reusable Python engine for clinical-note management, annotation, preprocessing, NLP inference, and LLM-powered research workflows.

My contribution

I designed and developed NLPMed-Portal and NLPMed-Engine, including longitudinal-note organization, annotation and adjudication, text preprocessing, model inference, and external LLM integration. I also packaged and released both NLPMed-Portal-Lite and NLPMed-Engine as independently installable open-source software.

  • Clinical NLP
  • Django
  • NLP Pipelines
  • SpaCy
  • LLM Integration
Portal Lite sourcePortal Lite on PyPIEngine sourceEngine on PyPI
Health-DeID review interface showing detected PHI and reviewer controls for a clinical record.

Health-DeID

Local research software for auditable de-identification of clinical text and selected structured fields through a user interface, command-line tools, and a Python API.

My contribution

I designed and developed the SQLite-backed workflow for PHI detection, category-specific transformations, automated validation, human review, surrogate generation, recoverable and revisable runs, cost tracking, and controlled exports across local and AWS-backed processing.

  • Clinical NLP
  • De-identification
  • Python
  • LLM Validation
  • Privacy
Source codePyPI package
NLPMed demo showing example input text and predicted labels from VTE-BERT and Bleed-BERT.

VTE-BERT & Bleed-BERT Models

Transformer-based clinical NLP models for VTE phenotyping and clinically relevant bleeding detection in longitudinal clinical notes, including public BERT and ModernBERT VTE classifiers fine-tuned on de-identified text.

My contribution

I developed the preprocessing, model-training, inference, evaluation, and patient-level aggregation workflows for VTE-BERT and Bleed-BERT, including VTE-BERT's internal and external validation pipeline. I later fine-tuned and released VTE-BERT-DeID and VTE-ModernBERT-DeID using Health-DeID and NLPMed-Engine.

  • Clinical NLP
  • BERT
  • ModernBERT
  • Transformers
  • Clinical Phenotyping
VTE-BERT-DeID on Hugging FaceVTE-ModernBERT-DeID on Hugging FaceInteractive demosVTE-BERT publication
Database-agent demo showing a natural-language cohort question and tabular results with yearly patient counts for pulmonary embolism and deep vein thrombosis.

Cancer Research Data Platform & Database Agent

A harmonized cancer-research data warehouse paired with an LLM agent for natural-language cohort queries and interactive data exploration.

My contribution

I designed the PostgreSQL warehouse and ETL pipelines, built large oncology cohorts from institutional and nationwide EHR data, and developed a LangGraph agent that selects relevant tables, generates and executes SQL, diagnoses errors, and summarizes results.

  • PostgreSQL
  • ETL
  • LangGraph
  • SQL
  • EHR Data
Database schemaInteractive ERD
Composite view of the decision-support application showing EHR-populated risk-factor selections and the resulting thrombosis.

Cancer-Associated Thrombosis Decision Support

A clinical decision-support application combining established thrombosis risk scores, bleeding exclusions, and guideline-based recommendations for oncology care.

My contribution

I developed Epic Caboodle SQL workflows to extract patient variables and generate PHI-free parameterized links. I also built the Next.js interface that prepopulates risk and bleeding-exclusion inputs and presents the resulting scores and guideline-based recommendations.

  • Next.js
  • SQL
  • Epic Caboodle
  • Clinical Decision Support
Live application
Federated-learning architecture showing three institutional sites exchanging model updates with an AWS-hosted NVIDIA FLARE aggregator while clinical data remains local.

Privacy-Preserving Federated Learning

A three-site federated-learning pilot enabling collaborative clinical machine-learning experiments across institutions while keeping source data local and exchanging only model updates.

My contribution

I designed and deployed the three-site NVIDIA FLARE architecture, including secure model-update exchange, AWS-hosted aggregation, MLflow experiment tracking, deployment automation, and a repeatable onboarding workflow for participating institutional sites.

  • NVIDIA FLARE
  • Federated Learning
  • AWS
  • MLflow
  • Distributed Systems

Earlier work

Earlier research

Before moving into clinical and healthcare applications, my research focused on algorithms and data structures for efficient similarity search.

Efficient Similarity Search in High-Dimensional Data

During my doctoral research, I developed machine-learning-based, cache-conscious, and tree-based indexing methods for efficient similarity search in high-dimensional data.

Doctoral dissertation