Machine Learning Researcher
Dr. Cescon Research Group, University of Houston
Houston, TX · –
Summer undergraduate researcher on multivariable glucose forecasting for patients with Type 1 diabetes, under Dr. Marzia Cescon.
- Built a gradient-boosted exercise detector from wearable sensor data that caught 98.5% of exercise episodes within 30 minutes at 0.71 false alarms per day.
- Led a multi-seed replication study in Python showing the headline results shrank to 20–50% of their reported size, which changed how the team evaluates models.
- Designed and trained a dual-branch CNN-LSTM in TensorFlow/Keras and fixed gate saturation with a variance-floor regularizer (9 of 9 fits trained stably).
- Found that raw wearable sensor signals directionally outperformed hand-engineered composite biomarkers for glucose forecasting across all 67 subjects.
Skills gained
- Python
- TensorFlow
- Keras
- scikit-learn
- Gradient boosting
- CNN
- LSTM
- Time-series forecasting
- Experimental design
- Reproducibility