Projects
Category

Deep Learning

TensorFlow projects across time-series forecasting, NLP transfer learning, and computer vision. Fewer notebooks, more end-to-end evals.

01

BitPredict

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. It progressed from a naïve sequential CNN baseline through Conv1D, LSTM, and multivariate Conv1D to a from-scratch replication of the N-BEATS architecture (Oreshkin et al., "Neural Basis Expansion Analysis for Interpretable Time Series Forecasting"), and finally a stacked ensemble of all of the above. The final ensemble scored MAE 566.77, RMSE 1072.96, MAPE 2.55%, and MASE 0.996, and beat every individual model. The honest writeup is that 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.

TensorFlowN-BEATSLSTMConv1DEnsembles

View on GitHub →

02

SkimLit

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. It moved from an sklearn baseline (72% accuracy) through Conv1D (78.6%), Universal Sentence Encoder transfer learning (71%), and token plus character embeddings (73.4%) to a 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.

TensorFlowTransfer LearningUSEEmbeddingsConv1D

View on GitHub →

03

Food Vision

Food101 classifier. Baseline, data augmentation, fine-tuning, hitting ~77% validation accuracy.

Computer vision classifier over the Food101 dataset, inspired by the original Food101 paper. It started with a simple CNN baseline that overfit quickly, added data augmentation (variance up, raw accuracy down, but the model generalizes better), moved to EfficientNet feature extraction, and finished by unfreezing the top layers for fine-tuning. Final validation accuracy was about 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.

TensorFlowEfficientNetTransfer LearningData AugmentationTensorBoard

View on GitHub →