Viraat Nellutla

My first convolutional neural network—built with Keras Sequential (conv/pool/dense). Accuracy? Low. Value? Huge. I learned how filters see, why stacking them explodes compute, and how training dynamics behave beyond the tutorial.

Lesson: “just add layers” isn’t a strategy. Tuning is.

Built a deeper CNN (16→32→64→128 filters) with BN + dropout, trained with Adam over 50 epochs on 50k images. Added a confusion-matrix view to spot where the model fumbles classes (and why).

Impact over accuracy: I learned to diagnose models, not just train them.

A scikit-learn pipeline using logistic regression + baselines on the Wisconsin dataset (569 samples, 30+ features). ~90% accuracy with clear evaluation (precision/recall/ROC). Designed for interpretability and next-step research.

Why it matters: This is the model that means the most to me.

Research-focused analysis of biomarkers and their correlation with diagnostic outcomes. Building feature-selection + interpretability first to guide clinical usefulness.

Goal: evidence, not just accuracy numbers.

LSTMs and BERT for tweet sentiment with a goal of longer-horizon trend signals. Beyond returns, I’m focusing on robust evaluation and avoiding hindsight bias.

Humbling truth: markets don’t care about your loss curve.