Projects.
Case studies with the decisions, trade-offs, and dead ends left in — the parts most portfolios delete.
Multimodal Breast Cancer Detection.
// b.sc. thesis · 2025Mammography alone misses context a clinician has on paper. The thesis asks: does fusing the image with tabular clinical features beat either signal alone — on a small, imbalanced, real dataset?
pytorch · medclip · tabnet · early fusionimg (medclip) ──┐
├─▶ concat ─▶ head ─▶ dx
tab (tabnet) ──┘
87.1% ROC-AUC, 82.7% accuracy — above both unimodal baselines.
Engineering Decisions & What Didn't Work.
// four dead ends, kept deliberately public — expand any entry
The obvious architecture — train each modality separately, average the outputs — capped out early. The modalities never got to interact, so cases where the image and the chart disagree were exactly the ones it failed.
→ shipped: single fusion head over concatenated embeddingsA few hundred patients cannot re-teach a foundation model. Unfreezing the full encoder memorized the training set while validation collapsed — the fix was accepting the pretrained features as-is.
→ shipped: frozen encoder; only the fusion head and tabnet trainEarly runs looked great until the confusion matrix told the truth. Model selection moved to ROC-AUC, training got a class-weighted loss, and every later comparison reports both.
→ shipped: weighted loss + auc-based model selectionTrees don't emit embeddings, so XGBoost couldn't participate in fusion. TabNet's learned representations concatenated cleanly with the image features — the fused model outperformed any tree pipeline.
→ shipped: tabnet embeddings as the tabular arm of the fusionWork history.
// two internships, outcomes onlyML Engineering Intern — Kontemporary Konsulting
- weekly consulting reports automated end-to-end, data collection to formatted memo
- corporate website designed, built, and shipped
Data & Automation Intern — TETFund
- document-processing workflows automated across departments
- validation scripts caught entry errors before they reached official records