Applied Machine Learning

Deep-learning projects, plus development and medical ML in progress

Machine learning projects

Face mask image classification (PyTorch)

A convolutional neural network in PyTorch that classifies face images as with mask or without mask. A supervised binary image-classification project covering the full pipeline: image preprocessing and augmentation, CNN training, and evaluation on held-out data. Built in Python.

X-ray fracture screening (TensorFlow)

Working with X-ray scan images to separate fracture from no fracture, using TensorFlow. The project pairs supervised labelling with unsupervised clustering of the scans to surface structure in the images and support annotation. Built in Python.

Work in progress

The projects below are works in progress. Several are exploratory and are currently paused or shelved for lack of funding; they are listed to show direction rather than finished results.

Poverty estimation from satellite imagery

Training convolutional neural networks on satellite imagery, calibrated against household-survey data, to produce small-area estimates of poverty and wealth indicators where survey coverage is sparse or infrequent. The aim is timely, low-cost poverty maps that complement traditional survey programs.

School and hospital mapping

Detecting and mapping schools and health facilities from satellite imagery to support service-delivery planning and equitable resource allocation, particularly where official facility registries are incomplete or outdated.

Informal cross-border trade

Informal cross-border trade is a large but poorly measured share of intra-African commerce. This work explores model-based estimation strategies that combine limited field observation with auxiliary data, a problem where careful survey design and statistical modeling matter as much as the data collection.

Offline reinforcement learning for sepsis treatment

Using the MIMIC-IV critical-care database to study offline (batch) reinforcement learning for sepsis treatment policies: learning from retrospective ICU trajectories, with emphasis on reliable off-policy evaluation so that learned policies can be assessed honestly before any claim of clinical relevance.


A common thread with my methodological research: when data are expensive (surveys, field studies, clinical records), experimental design and sample-efficient methods decide what is learnable.