MBBS and machine learning engineer building end-to-end clinical systems across time-series modelling and clinical NLP

Currently contributing to ML research workflows at radnomics ltd involving large-scale data augmentation, unsupervised radiology report generation, and proprietary LLM evaluation

Python · PyTorch · Hugging Face Transformers · Scikit-learn · Pandas · FastAPI · Docker · Google Cloud Run · GitHub Actions

• Hybrid clinical NLP system generating structured entity outputs from adult ICU progress notes
• Implemented rule-based regex extraction schemas for recall-focused extraction of 3 clinical entity types
• Fine-tuned a BioClinicalBERT classifier on 1000+ manually annotated entities, and integrated precision-oriented threshold tuning for final entity validation
• Extracted 780,000+ structured entities from a preprocessed ICU corpus of 160,000+ notes (30,000+ stays)
• Transformer validation achieved +45.9% in precision and −83.3% in false positives relative to rule-only baseline
• Deployed inference pipeline as stateless, containerised API on Google Cloud Run with GitHub Actions versioning

Live API · Repository · Zenodo DOI

Python · PyTorch · LightGBM · Scikit-learn · Pandas · NumPy · SHAP

• Dual-architecture ICU early warning system combining a Temporal CNN (TCN) and LightGBM to predict NEWS2-derived deterioration outcomes across 3 clinical risk dimensions
• Transformed clinical data across 140 ICU stays using CO2 retainer logic, GCS mapping, and oxygen protocols
• Engineered 171 timestamp-level features (8 vital parameters; 96-hour windows) and 40 aggregated patient-level features from 70,000+ extracted time-series observations
• TCN achieved +9.3% AUC improvement for acute-event detection; LightGBM achieved −68% Brier score and −48% RMSE for prolonged risk exposure
• Implemented clinician-interpretable SHAP and saliency mapping for feature contribution insights

Repository · Zenodo DOI

Applied Machine Learning Engineer @ RadNomics Ltd

• Processed 2.3M+ radiology reports and developed a data augmentation pipeline generating 17M+ report pairs across 7 clinical reconstruction tasks
• Ran large-scale ML research workflows within containerised remote environments using Git-based version control and distributed GKE-based cloud infrastructure
• Built a proprietary LLM benchmarking framework evaluating 6 frontier language models across 42,000 reconstructions, quantifying performance using lexical, semantic, and clinical entity-aware metrics
• Translated performance insights into production system workflows, directly optimising downstream efficiency of report generation, editing, and dictation features
• Currently evaluatung proprietary open models fro baseline comparison

• Machine Learning: PyTorch, TensorFlow/Keras, Scikit-learn, LightGBM, Hugging Face Transformers, Clinical NLP, LLM Evaluation
• DevOps: Google Cloud Platform (GKE, Cloud Run), Kubernetes, Docker, FastAPI, GitHub Actions (CI/CD)
• Data & Engineering: Python, Pandas, NumPy, SQL (PostgreSQL/MySQL), Seaborn

• MSc, Computer Science with Artificial Intelligence @ City St George’s, University of London
• MBBS, Medicine @ Norwich Medical School, University of East Anglia

• Lacertus syndrome and its surgical management using WALANT - our first 12 cases (Research Poster)
• Giant trichoblastic carcinoma initially misdiagnosed as basal cell carcinoma (Case Report)

• Head and Neck Surgery, Integrated Care Pathway Surgical Proforma Audit
• Plastic Surgery, Free Flap Surgical Outcomes Audit

• Clinical Informatics: EHR Systems (ICE, SystmOne, MediViewer, EPMA), NEWS2, GDPR
• Clinical Research: Audit Methodology, Literature Review, Critical Appraisal, Manuscript Preparation

• Yip, S. (2026). Clinical Entity Extraction-Validation System (1.0.0). Zenodo. https://doi.org/10.5281/zenodo.20018309
• Yip, S. (2026). Time-Series ICU Patient Deterioration Predictor (1.0.0). Zenodo. https://doi.org/10.5281/zenodo.18487174