• Matryoshka Representation Learning for Medical Coding

  How nesting embeddings at multiple dimensions solves a real inference problem in clinical NLP — and why the geometric intuition matters more than the loss function.

• Feb2025 - Aug2025
  Datalogic USA, Inc.
  ML Research Intern
  • Probabilistic Modeling & Latent Space Geometry: Architected a custom von Mises-Fisher Mixture Model (vMFMM) via Expectation-Maximization to cluster 128-dimensional hyperspherical embeddings from ArcFace/CCE losses — 98% Top-4 Macro recall, 85% Top-1 Macro, a 1–5% improvement over Euclidean GMM baselines across 6 architectures on 500K+ images.
  • Data Pipeline & Incremental Learning: Engineered a 200-day continuous learning simulation benchmarking 6 ML architectures (XGBoost, SVM, Random Forest, etc.); demonstrated peak accuracy within a 50-day window, directly informing deployment timelines and data collection strategy.
  • Hyperparameter Ablation: Executed ablation studies across 380K+ training samples; statistically validated a 5-component vMFMM as optimal for long-tail imbalanced class distributions.
  • Robustness & OOD Testing: Stress-tested the vision pipeline with targeted label corruption and Gaussian feature noise; validated zero-shot OOD generalization under sensor degradation conditions.

• AstraGraph

  Full-stack GraphRAG system combining a Neo4j property graph and Qdrant vector store to answer structural and semantic queries over Python codebases.
• Matryoshka-ICD

  Multi-label ICD coding for clinical radiology reports using BioClinical-ModernBERT and Matryoshka Representation Learning — adaptive embeddings across dimensions without retraining.
• vMFMM: Hyperspherical Clustering for Production Vision

  Custom von Mises-Fisher Mixture Model replacing Euclidean GMM in a production edge vision pipeline — 98% Top-4 Macro recall across 500K+ images from 6 encoder architectures.

Let's Connect

• github /
• linkedin /
• huggingface /
• hassensaid32@gmail.com