Deep Learning
TensorFlow projects across time-series forecasting, NLP transfer learning, and computer vision. Fewer notebooks, more end-to-end evals.
BitPredict: N-BEATS time-series ensemble
Walked the model zoo end-to-end, landed on an ensemble with MAPE 2.55% on Bitcoin price.
A TensorFlow time-series forecasting project on Bitcoin price. Progressed from a naïve sequential CNN baseline → Conv1D → LSTM → multivariate Conv1D → a from-scratch replication of the N-BEATS architecture (Oreshkin et al., "Neural Basis Expansion Analysis for Interpretable Time Series Forecasting") → a stacked ensemble of all of the above. Final ensemble: MAE 566.77, RMSE 1072.96, MAPE 2.55%, MASE 0.996. The ensemble beat every individual model. Honest writeup: crypto prices depend too much on outside factors for any forecasting model to truly "win", but the experiments mapped the trade-offs between architectures cleanly.
SkimLit: NLP medical-text skimmer
Replicated the PubMed 200k RCT paper. Token, char, and positional embeddings hit 83% accuracy.
A TensorFlow NLP classifier that labels the sections of medical RCT abstracts (Background / Objective / Methods / Results / Conclusions) so long papers can be skimmed. Replicates the PubMed 200k RCT paper. Progression: sklearn baseline (72% accuracy) → Conv1D (78.6%) → Universal Sentence Encoder transfer learning (71%) → token + character embeddings (73.4%) → final model combining token, character, and positional embeddings (83.2% accuracy, F1 0.83), matching the original paper's ballpark. A good case study in why embedding composition matters more than model size.
Food Vision: EfficientNet transfer learning
Food101 classifier. Baseline, data augmentation, fine-tuning, hitting ~77% validation accuracy.
Computer vision classifier over the Food101 dataset, inspired by the original Food101 paper. Workflow: simple CNN baseline (overfit quickly) → add data augmentation (variance up, raw accuracy down, but the model generalizes better) → EfficientNet feature extraction → unfreeze top layers for fine-tuning. Final validation accuracy ~77%, with TensorBoard curves and confusion matrices tracking every step. Also tried EfficientNetB7 and documented that bigger wasn't better for this dataset. A useful "simple > complex" lesson captured in the writeup.