Past
Research.

Chest X-ray datasets are large and costly to label, creating demand for automated classifiers that perform reliably under real-world class imbalance.

Trained and compared a 100-tree random forest and a custom 3-layer CNN on ~6,000 labeled images, using stratified splits and class weighting to handle a 73/27 imbalance.

The random forest reached 93% accuracy (F1 0.91), outperforming the CNN (86%, F1 0.89) while training faster — establishing a strong, cost-effective baseline for clinical screening.

Fraud detection systems must catch rare transactions in heavily imbalanced datasets without overwhelming analysts with false positives — a challenge that off-the-shelf accuracy metrics obscure.

Tested logistic regression, clustering methods, and random forest on 1 million synthetic transactions, evaluating precision, recall, and feature importance alongside overall accuracy.

Random forest achieved ~95% accuracy and identified transaction amount, account age, and timing as the strongest fraud signals. Low recall across all models confirmed that class imbalance requires targeted mitigation beyond model selection alone.

As AI automates entry-level roles, new graduates face a narrowing job market and the traditional career pipeline — from degree to entry-level work to expertise — is breaking down.

Analyzed AI's displacement effects across sectors including healthcare and finance, weighing efficiency gains against opportunity costs for students, universities, and the broader labor market.

Recommends deploying AI as a human amplifier rather than a replacement — preserving career entry points while capturing efficiency gains — and argues that productivity and opportunity are not mutually exclusive if the deployment model is designed with both in mind.

The original paper's dataset was not publicly available, making it impossible to reproduce its findings and evaluate whether the conclusions generalize across different data distributions.

Built a new dataset using CWUR rankings with 80 simulated users and 1,600 ratings. Replicated SVD, KNN Basic, and KNN with Baselines via 5-fold cross-validation, then extended the study with MAE, R², a neural network model, hyperparameter tuning, and a second evaluation on a real-world video game ratings dataset.

KNN with Baselines outperformed SVD on the simulated data, while SVD won on real-world data — demonstrating that recommender performance is heavily dataset-dependent and that replication without the original data demands careful interpretation of divergent results.

Social media algorithms optimize for engagement, creating echo chambers where users see increasingly narrow content — reinforcing existing beliefs and limiting exposure to different perspectives without users realizing it.

Interviewed regular social media users, reviewed research on popularity bias and algorithmic fairness, and designed "Perspective Balance" — a feature that detects repetitive feed patterns and surfaces contextually explained "branch out" posts.

A concrete, platform-agnostic feature design that retains personalization while adding diversity, transparency, and user control — addressing fairness and manipulation concerns without removing the algorithmic recommendations users actually value.

Fixed temporal patching — the standard tokenization strategy for ECG Transformers — splits heartbeat structures across token boundaries, treating cardiac recordings as generic time series rather than structured physiological signals.

Compared fixed patches against three beat-aligned tokenizers (resampled beats, adaptive pooled beats, resampled beats + R-R interval timing) across 12-lead diagnostic classification on PTB-XL and 60-second rhythm classification on Icentia11k, using matched Transformer architectures and SSL objectives.

Resampled beat tokens matched the best fixed-patch performance on PTB-XL while using 9× fewer tokens (11.2 vs 100), and showed greater run-to-run stability on imbalanced rhythm data — establishing beat-synchronous tokenization as a compact, competitive alternative with a meaningful inductive bias for cardiac modeling.