PhD · Data Scientist · NLP Researcher

Building systems
at the edge of
language & data.

I design production ML pipelines and publish research on LLM uncertainty, multi-label classification, and NLP — bridging academic rigor with real-world deployment.

View Research Explore Projects
12+
Publications
90
Citations
PhD
Industrial & Systems Eng.
5+
Models Shipped
01

About Me

I'm an Data Scientist at Cooper University Health Care and an independent researcher specializing in large language models, uncertainty quantification, and NLP for clinical and financial domains.

My PhD in Industrial and Systems Engineering and Master's in Finance from Paris Dauphine give me an unusual vantage point: I think about systems, uncertainty, and optimization simultaneously across both technical and economic layers.

In production I've built a hospital-policies RAG chatbot (ChromaDB + Azure OpenAI), a call-mining quality pipeline processing thousands of recordings monthly, and a suite of forecasting tools. In research I treat LLM stochasticity as signal, not noise.

LLM Uncertainty Healthcare NLP Multi-label Classification RAG Systems Conformal Prediction Financial ML Azure OpenAI Python
Hajar Sakai
Hajar Sakai

Education

  • PhD · Industrial & Systems Engineering
  • MSc Finance · Paris Dauphine

Current Role

  • Data Scientist · Cooper University Health Care
  • Independent ML Researcher

Research Interests

  • LLM stochasticity as latent signal
  • Lookahead bias in financial LLMs
  • Bayesian NLP for clinical text
  • Multi-label text classification
02

Selected Research

01
JMIR AI · 2026
LLMs for Healthcare Text Classification: A Systematic Review
Comprehensive systematic review mapping the landscape of large language model applications in clinical and healthcare text classification, synthesizing evidence across benchmarks, architectures, and deployment contexts.
Published
02
IISE 2026 · Conference
Bayesian Latent Variable Modeling of LLM Stochasticity
A novel framing that treats temperature-induced variance in LLM outputs as a structured latent signal rather than noise, integrated with conformal prediction for calibrated uncertainty bounds.
Forthcoming
03
Working Paper
Lookahead Bias in Financial LLMs
Investigating temporal leakage in LLM-based financial forecasting — identifying conditions under which training data leads to systematically optimistic backtests and proposing evaluation protocols to surface hidden bias.
In Progress
04
Multi-venue · 2023–2025
QUAD-LLM-MLTC / KDH-MLTC / HAMLET Frameworks
A family of multi-label text classification frameworks combining ensemble LLM prompting, knowledge distillation, and hierarchical label structures — demonstrated on clinical and biomedical corpora.
12+ Papers
03

Industry Projects

Production · Healthcare
Policies RAG Chatbot
End-to-end retrieval-augmented generation system over 178 hospital policy documents. Features multi-intent routing (semantic / index / clinical-links paths), delta indexing, inline citations, and Azure OpenAI embeddings.
ChromaDB Azure OpenAI Streamlit Python
Production · Analytics
Call Mining & Quality Pipeline
Automated pipeline processing thousands of monthly call recordings via Azure Blob Storage and OpenAI Whisper transcription. Stratified sampling and LLM scoring across 9 quality dimensions for the Patient Access Center.
Azure Blob gpt-audio LLM Scoring Fabric
Finance · NLP
10-K Sentiment Analyzer
Technical assessment combining Loughran-McDonald financial lexicon with LLM-based sentiment analysis on 10-K filings. Designed to surface tone shifts predictive of firm-level risk across reporting periods.
LM Lexicon LLM NLP Python
ML · Forecasting
Provider Star Rating Model
End-to-end ML workflow for provider performance prediction and star rating at Cooper. Combines structured EHR features with NLP signals from patient feedback, with statistical significance testing throughout.
Scikit-learn NLP Statistics Sentiment Analysis
Let's connect

Open to research
collaborations & roles.

I'm interested in academic positions, research collaborations, and senior ML/NLP industry roles at the intersection of language models and high-stakes domains.

Email GitHub LinkedIn Google Scholar Download CV