I'm a Computer Vision / ML engineer with a focus on healthcare applications — systems where getting the prediction wrong has real consequences for real people.

My work sits at the intersection of:

• Clinical measurement — translating domain knowledge (Clarke EGA, ISO 15197, MARD) into loss functions and evaluation protocols
• Lightweight deployment — getting models onto Apple Watch, mobile, and edge devices
• Honest evaluation — I care about clinical metrics, not just RMSE on a validation split

Currently exploring MSc programs in Medical AI / ML (2025 intake). Looking for research roles at the intersection of ML and healthcare.

Predicting blood glucose 30 and 60 minutes ahead from continuous CGM sensor data.
LSTM + Temporal CNN trained on OhioT1DM, evaluated with Clarke Error Grid.
Exported to Core ML (197 KB) and deployed on watchOS via HealthKit.

PyTorch Core ML Swift watchOS HealthKit Streamlit

Live demo: glucosense-ai.hf.space

From-scratch replication of Ronneberger et al. (2015) on the DRIVE fundus dataset.
Includes architecture ablation (BN, dropout, depth) and Grad-CAM visualization.

PyTorch Albumentations DRIVE dataset Grad-CAM

Bilateral facial symmetry analysis using MediaPipe Face Mesh (468 3D landmarks).
Measures per-region geometric asymmetry normalized to inter-ocular distance.
Includes explicit ethical limitations section — this measures geometry, not attractiveness.

MediaPipe OpenCV Gradio HuggingFace Spaces

Live demo: face-analysis-app.hf.space

Working notes on the math and code behind modern ML — linear algebra, attention, CNNs, paper notes.
Implemented from scratch: multi-head attention, dilated convolutions, Grad-CAM, LoRA intuition.

Jupyter PyTorch NumPy Matplotlib

Deep Learning    PyTorch · Core ML · TorchScript · ONNX
CV / Medical     Segmentation · Detection · Time Series · CGM Analysis
Deployment       watchOS · HuggingFace Spaces · Streamlit · Gradio
Languages        Python · Swift · SQL
Infrastructure   Git · Docker · W&B

• Clinical AI evaluation — gap between academic benchmarks and clinical utility
• Uncertainty quantification — models should express doubt, not be confidently wrong
• Efficient inference — making research-grade models deployable on wearables and mobile
• Multimodal medical data — fusing CGM + insulin + activity for glucose prediction

• Isensee et al. (2021) — nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation
• Dao et al. (2022) — FlashAttention: Fast and Memory-Efficient Exact Attention
• Hu et al. (2022) — LoRA: Low-Rank Adaptation of Large Language Models