CARDIAC-3D · SIM n=01 · v 4.6.2
Clinical-AI Environment Initializing
ECG Lead II Signal Locked · 72 BPM
Cardiac-3D Runtime Systole Ready
Patient Model n=01 Loaded
Session Initialized
Available · Research Engineer & Pre-doctoral Fellow · Fall 2026

Building clinical AI
for the real world.

MS Data Science (AI) at Northwestern, August 2026. Clinical-AI researcher and developer working at the intersection of EHR data, large language models, and decision support. Published in cardiovascular AI. Previously at Shirley Ryan AbilityLab.

See the work → Get in touch
Spec · CARDIAC-3D Sim n=01 Live
CO5.4L/min
SV75mL
MAP92mmHg
Lead II 72 bpm NSR
♥ HR72bpm SpO₂98% MAP92mmHg eGFR96mL/min Hb14.2g/dL Na140mEq/L K4.1mEq/L Cr0.9mg/dL Lac1.1mmol/L
INPUT · HIDDEN · OUTPUT
01  /  About

Clinical AI you can actually trust.

I build clinical-AI tools clinicians actually use. My work lives in the messy middle of healthcare AI, where EHR data, validated phenotypes, and rigorous LLM evaluation meet. Pipelines before models. Validation before novelty. Research that ships.

01

Clinical research

Active research-assistant role at Northwestern Medicine. EHR phenotyping pipelines for federated multi-site cohorts under IRB.

02

NLP & LLM evaluation

Hallucination detection, evidence-citation patterns, and prompting comparisons on frontier models in clinical text.

03

Full-stack delivery

Python, FastAPI, Next.js, Streamlit. I ship deployed systems with versioning and feedback, not slide decks.

04

Anthropic-native

Multiple production projects on the Anthropic SDK. Prompt engineering, evaluation harness design, agentic workflows.

AT GC TA CG AT GC TA CG
02  /  Current Research

EHR phenotyping at scale.

InstitutionNorthwestern Feinberg School of Medicine
DepartmentDivision of Rheumatology
PIDr. Rosalind Ramsey-Goldman, MD DrPH
Co-IDr. Catherine Chen, MD MPH
ProjectClinical EHR research, IRB
TimelineFeb 2026 — Present
StatusActive  ·  pre-publication

Reproducible Python pipelines for clinical cohort identification on federated EHR data.

Working under a senior principal investigator on migrating legacy R research code into maintainable, version-controlled Python workflows. Focus on reproducibility, cohort validation, and cross-language parity across heterogeneous EHR systems. Current work is under active IRB protocol. Detailed methodology, sample sizes, and findings are pre-publication and available on request.

Source
Federated EHR
→
Engineering
Python · SQL · pandas
→
Output
Validated cohort
Python SQL pandas ICD-10 SNOMED CT CPT Federated EHR Reproducible research
Live · ICD-10 · SNOMED
●ICD-10M32.10SLE ●SNOMED55342001neoplasm ●CPT99213visit · est ●cohortVALID ●phenotype.match0.94 ●ICD-10I50.21heart failure ●RxNorm19651aspirin ●LOINC2160-0creatinine ●flagREVIEW ●ICD-10N18.3CKD III ●SNOMED42343007hypertension ●validationPASS ●ICD-10M32.10SLE ●SNOMED55342001neoplasm ●CPT99213visit · est ●cohortVALID ●phenotype.match0.94 ●ICD-10I50.21heart failure ●RxNorm19651aspirin ●LOINC2160-0creatinine ●flagREVIEW ●ICD-10N18.3CKD III ●SNOMED42343007hypertension ●validationPASS
03  /  Projects

Work that shipped.

A.01  /  Deployed · Live

A 10-module clinical decision-support portal.

Ten independent clinical modules under a unified Streamlit shell, each wrapping a trained model with a clinician-facing interface: PDF lab uploads, image upload for chest X-rays, conversational interpretation through an AI clinical assistant on the Anthropic Messages API. Continuous deployment, model versioning, evaluation harness.

StackPython · TensorFlow · scikit-learn · Streamlit · Anthropic SDK
Headline1D-CNN for ECG · MIT-BIH benchmark (held-out splits)
StatusLive in production
Open live demo ↗
01ECG Rhythm Classification1D-CNN · Held-out
02Chest X-Ray TriageMobileNetV2
03Heart Disease RiskRF · 83%
04CBC InterpretationRule + ML
05Diabetes ScreeningClassifier
06Lipid Panel AnalysisRisk Score
07Kidney FunctioneGFR Model
08Lab Report UploadOCR + LLM
09AI Clinical AssistantAnthropic API
10Multi-Modal FusionEnsemble
A · Problem
Most research ML never leaves the notebook.

Clinicians and patients need interpretable, fast, point-of-care decision support across heterogeneous lab and imaging data. I wanted to prove research code could become a real, deployable tool.

B · Architecture
10 modules under one shell.

Independent modules sharing a Streamlit app, unified theming, modular evaluation harness. Continuous deployment on Streamlit Cloud with model versioning and rollback. Shared session state.

C · Models
From 1D-CNNs to calibrated RFs.

ECG rhythm classification (custom 1D-CNN, held-out test) on the MIT-BIH arrhythmia set. Chest X-ray triage via MobileNetV2 transfer learning. Tabular modules using calibrated classifiers with rule-based safety guards.

D · Differentiator
An AI clinical assistant module.

Built on the Anthropic Messages API, it accepts PDF and image uploads, holds multi-turn session memory, surfaces safety disclaimers, and is cross-module aware. The same architecture scaffolds any clinical decision-support system.

70 60 50 PROMIS · PI · 12mo MIC ±5
ShippedMay 2026

PROMIS Clinical Dashboard

Six-domain PROMIS T-score estimator from clinical text. IRT-calibrated scoring, MIC-flagged change detection, synthetic-data demo. Local-first design, no PHI required.

StreamlitPythonscikit-learnPlotlyGroq
Open project↗
Patient #017 · 18-month trajectory RADIOLOGY ECOG · PERFORMANCE ⚡ INFLECTION
ShippedApr–May 2026

ONCOTRAJ · Oncology Toolkit

End-to-end clinical-NLP replication of three published methods on a real 30-patient synthetic oncology cohort. Trajectory smoothing, performance-status extraction, inflection-point detection across radiology + ECOG signals. Live API.

Python FastAPI SQLite SQLAlchemy Anthropic SDK
Open project↗
MedReason-Bench · n=6 models Opus 4.7 Sonnet 4.6 GPT-5 Gemini 2.5 Llama 4 Med-PaLM 3 TBD TBD TBD TBD TBD TBD Composite · accuracy × calibration × hallucination
In Development · Mar 2026 — Present

MedReason-Bench

Open benchmark for clinical reasoning in LLMs across cardiology and rheumatology splits. Evaluation harness in development — methodology, prompt strategies, and dataset construction documented; full eval results pending Q3 2026 release.

Python Anthropic SDK OpenAI SDK Google AI SDK pytest
Open project↗
Live · Eval Stream
●opus-4.7pending ●sonnet-4.6pending ●hallucinationpending ●calibrationECE pending ●promptdouble-filter ●cohortvalidated ●imputationMICE-5 ●test.splitheld-out ●driftmonitored ●PROMIST-score ●MIC±5 ●eval.harnessCI ●opus-4.7pending ●sonnet-4.6pending ●hallucinationpending ●calibrationECE pending ●promptdouble-filter ●cohortvalidated ●imputationMICE-5 ●test.splitheld-out ●driftmonitored ●PROMIST-score ●MIC±5 ●eval.harnessCI
LEAD II · 25 mm/s · 10 mm/mV
04  /  Publications

Peer-reviewed research.

Published 2024 · Systematic Review

Edge-AI Meets the Heart: Real-Time Cardiovascular Monitoring with Cloud-Connected Wearables

Sami AA, Khan MK, Kumar S, Shiwlani A.
Advances in AI & Machine Learning · Scopus · Web of Science

A systematic review evaluating Edge AI in wearable devices for real-time cardiovascular monitoring, alongside cloud-based systems for predictive analytics and personalized therapy. Literature search across Google Scholar, PubMed, and Web of Science (2015–2025), 937 articles narrowed to 34 by relevance and quality. CNN and RNN models achieved 85–98% diagnostic accuracy on edge AI hardware. Federated learning addressed privacy; cloud integration enabled real-time processing.

Read full paper ↗ ORCID record ↗
937
Articles surveyed
34
Studies synthesized
98%
Peak CNN accuracy
05  /  Approach

How I think about clinical AI.

Great clinical AI isn't the model. It's the pipeline feeding the model.

01

Pipelines before models

The model is the easy part. What makes clinical AI reproducible is what feeds it: cohort validity, code-set reconciliation, validation across languages and sites. I spend my time where the scientific risk actually lives.

02

Generalist tools, specialist outcomes

Python, SQL, a notebook, a version-controlled pipeline. The technology should be boring and reliable so the clinical question can be ambitious. Maintainability over novelty.

03

Research, shipped

Code in a notebook is a hypothesis. Code that's deployed, versioned, and used is research. My work ships, from the Healthcare AI Portal to the pipelines I build for active IRB studies.

Live · Vitals
♥HR72bpm ●SBP118mmHg ●DBP76mmHg ●SpO₂98% ●RR14/min ●Temp37.0°C ●eGFR96mL/min ●Hb14.2g/dL ●WBC7.4k/μL ●PLT240k/μL ●Na140mEq/L ●K4.1mEq/L ●rhythmNSR ♥HR72bpm ●SBP118mmHg ●DBP76mmHg ●SpO₂98% ●RR14/min ●Temp37.0°C ●eGFR96mL/min ●Hb14.2g/dL ●WBC7.4k/μL ●PLT240k/μL ●Na140mEq/L ●K4.1mEq/L ●rhythmNSR
06  /  Experience

A trajectory through clinical research.

Feb 2026
— Present

Research Data Analyst

Northwestern Feinberg School of Medicine · Division of Rheumatology
PI · Dr. Rosalind Ramsey-Goldman, MD DrPH  ·  Co-I · Dr. Catherine Chen, MD MPH

Graduate research assistant building Python and SQL pipelines for clinical cohort identification on federated EHR data. Working under active IRB protocol on pre-publication research.

PythonSQLFederated EHRClinical research
2024
— Dec 2025

Research Data Analyst

Shirley Ryan AbilityLab · Cotton Lab
PI · Dr. R. James Cotton, MD PhD

Supported clinical AI research in motor rehabilitation. Built and maintained data pipelines for high-resolution motion-capture datasets across multi-camera rigs. Contributed to dataset curation, quality review, and feature engineering for downstream ML model development. Collaborated with clinicians, biomechanists, and ML engineers on documentation, reproducibility, and preprocessing routines feeding publishable research on wearable-sensor and vision-based clinical assessment.

Motion captureClinical AIPythonData pipelines
Aug 2025
— Present

Analytics Coordinator

Northwestern Data & Technology Student Leadership Council

Analytics Coordinator for DTSLC, a joint MSDS and MSIS governance body under SPS. Own end-to-end analytics and survey infrastructure: instrument design, Zoho Forms build-out, Bitly distribution, Eventbrite outreach to 1,000+ attendees, and post-event analytics. Led Spring 2026 annual feedback program and served as analytics lead for the April 2026 Healthcare AI Startup Event featuring Dr. Syamala Srinivasan and Frank Bernhard.

Survey designZohoEventbriteStakeholder reporting
May 2025
— Aug 2025

Data Analyst

NOW Foods

Data Analyst supporting enterprise supply chain and demand analytics. Built Tableau dashboards sourced from Oracle Demantra and ASCP, architecting reporting flows that drove a 22% increase in operational efficiency and a 15% improvement in forecast accuracy through statistical demand modeling and safety-stock optimization.

TableauSQLOracleForecasting
07  /  Stack

Tools I use daily.

Languages

  • Python
  • SQL
  • R
  • JavaScript / TypeScript
  • Shell

ML & Data Science

  • scikit-learn
  • XGBoost / LightGBM
  • TensorFlow / Keras
  • PyTorch
  • pandas · NumPy · SciPy

Clinical Data

  • EHR phenotyping
  • ICD-10 · SNOMED CT · CPT
  • UMLS · OMOP exposure
  • FHIR familiarity
  • REDCap · IRB · CITI

LLM & NLP

  • Anthropic SDK
  • Prompt engineering
  • Zero-shot · CoT · Few-shot
  • Evaluation harness design
  • Hallucination detection

Web & Apps

  • Streamlit
  • FastAPI
  • Next.js · React
  • Plotly · Recharts · D3
  • Tableau · Power BI

Databases

  • PostgreSQL
  • MySQL
  • MongoDB
  • Neo4j
  • SQLite · SQLAlchemy

Infra

  • Docker
  • Git · GitHub Actions
  • AWS
  • Streamlit Cloud
  • Netlify · Vercel

Methods

  • Cohort identification
  • Validation design
  • Statistical modeling
  • Transfer learning
  • CNNs · RNNs
08  /  Get in touch

Let's build something that matters.

Open to research-engineer, pre-doctoral fellow, and clinical-AI research assistant roles, clinical-AI collaborations, and conversations starting Fall 2026. Strongest fit: academic medical centers, clinical research labs, and health-AI groups working on EHR, imaging, or wearable data.

LinkedIn ↗ GitHub ↗ ORCID ↗ Publication ↗