Flagship stations. Archive when you need breadth.

Clinical·In progress

Patient enrollment analysis

A FastAPI dashboard that turns historical reprojection data into the operational view used to plan and triage active studies: LSE shift tracking, per-study enrollment velocity, TA rollups, and an admin layer for inclusions and exclusions.

Vision·Shipped

Leukemia classification

An EfficientNetB7 classifier with rotation, flip, and color-jitter augmentations. Accuracy lifted roughly thirteen points; held-out test set reached 98.9%. Became the technical base for the Berkeley MIDS “A World for Every Child” screening prototype.

98.9% held-out

Language·Shipped

Named entity recognition

A custom spaCy NER model extracting person, location, and organization entities from feedback text, wired as a feature-engineering step into downstream segmentation and CX pipelines.

95% P / 90% R

Predictive·Shipped

Customer churn prediction

Random Forest, Gradient Boosting, and a small neural net trained on the same cohort. Random Forest won on 92% accuracy and 0.85 F1 and became the retention team’s targeting list. Age, income, and occupation dominated feature importance.

0.85 F1

Language·Shipped

Customer sentiment model

A composite NLP stack — topic modeling, sentiment scoring, and keyword extraction — with a TensorFlow LSTM head, producing a single priority signal for ops.

Systems·In production

Cafe point-of-sale

A Flutter POS against Postgres with Square Mobile Payments, Apple Pay, and Zettle. Order entry with full drink customization, cash register with daily rollover and discrepancy tracking, shift attendance, and historical reporting. Paired later with self-service kiosk, dual-screen menu board, and a FastAPI receipt-printer service.