Mina Ekramnia

– Architected end-to-end production ML pipelines for automated document review and sensitive information classification for Education Global Practice and ITS AI/ML Solutions, processing thousands of scanned PDFs through OCR (Azure Document Intelligence), preprocessing, and document-based chunking. Implemented multi-level classification at document, paragraph, and token granularity using Python and LLM-based NLP (Vertex AI Gemini) with reinforcement learning and fine-tuning. – Built data pipelines in Databricks using Python, SQL, and Bash to extract, transform, and aggregate document metadata from Azure data warehouses. Developed evaluation frameworks using scikit-learn for classification metrics (precision, recall, F1) and FAISS for semantic similarity matching, achieving 60% ROI and increasing classification accuracy by 30% (VPU Team Award). – Engineered cost-optimized systems using intelligent prompt caching and workflow optimization, reducing API costs by 90% while maintaining accuracy across 9+ categories. Deployed models to production on Azure with automated monitoring and versioning (MLOps), integrated into Independent Evaluation Group workflows.

– Processed and analyzed PET and MRI brain images for early detection of Alzheimer's and Down syndrome using neuroimaging pipelines (FreeSurfer, FSL, SPM, ANTs) in MATLAB. Extracted structural and functional brain features, performed image segmentation and registration, and developed automated analysis workflows to identify biomarkers for neurodegenerative disease progression.

– Built predictive models to estimate disease progression rates using large-scale national surveillance datasets. Applied Bayesian inference, survival analysis, and causal inference methods (difference-in-differences, propensity score matching) in R and Stata, developing statistical pipelines to integrate multiple data sources and adjust for confounding variables. – Published findings in Epidemiology (link) and contributed to policy-relevant modeling, collaborating with Harvard Prevention Policy Modeling Lab, CDC, Stanford, and Yale.

– Architected end-to-end production ML pipelines for automated document review and sensitive information classification for Education Global Practice and ITS AI/ML Solutions, processing thousands of scanned PDFs through OCR (Azure Document Intelligence), preprocessing, and document-based chunking. Implemented multi-level classification at document, paragraph, and token granularity using Python and LLM-based NLP (Vertex AI Gemini) with reinforcement learning and fine-tuning. – Built data pipelines in Databricks using Python, SQL, and Bash to extract, transform, and aggregate document metadata from Azure data warehouses. Developed evaluation frameworks using scikit-learn for classification metrics (precision, recall, F1) and FAISS for semantic similarity matching, achieving 60% ROI and increasing classification accuracy by 30% (VPU Team Award). – Engineered cost-optimized systems using intelligent prompt caching and workflow optimization, reducing API costs by 90% while maintaining accuracy across 9+ categories. Deployed models to production on Azure with automated monitoring and versioning (MLOps), integrated into Independent Evaluation Group workflows.