Deep dive into TensorFlow’s multi-level APIs — mastering everything from raw ops to custom layers and high-level Keras models. Showcases building flexible pipelines, mixing abstraction levels, and crafting efficient, production-ready training workflows.

Transfer learning in TensorFlow/Keras using ImageNet-pretrained CNNs (e.g., VGG16, ResNet50) for CIFAR-10 classification. Demonstrates feature extraction, custom classifier heads, and fine-tuning workflows.

Loading, inspecting, and preprocessing datasets using TensorFlow Datasets (TFDS). Demonstrates train/test split handling, image normalization, batching, shuffling, caching, and visualization to build efficient, ready-to-train deep learning pipelines.

A series of Jupyter notebooks exploring TensorFlow and Keras APIs for deep learning. Covers CNNs on CIFAR10/MNIST, subclassing models, transfer learning, data pipelines, augmentation, image loading, low-level APIs, and custom layers/functions. Demonstrates practical implementations with fine-tuning, preprocessing, and visualization techniques.

Implementation of multiple deep learning models using Keras Functional API, including a CNN on MNIST, a multi-input/multi-output example, and a toy ResNet on CIFAR-10. Demonstrates advanced model architectures, tf.data pipelines, and practical training workflows with TensorFlow Datasets.

Hands-on deep dive into TensorFlow’s low-level API — covering tensor creation, shape manipulation, broadcasting, slicing, and math operations. Builds a strong foundation for custom deep learning workflows without relying solely on high-level abstractions.

Practical implementation of image data augmentation using TensorFlow and Keras preprocessing layers. Includes real-time transformations, visual comparisons, and training integration to improve model generalization and reduce overfitting without adding new data.

A hands-on guide to automatic differentiation in TensorFlow using tf.GradientTape. Covers computing gradients for variables vs. constants, using tape.watch(), visualizing derivatives, and handling multiple parameters.

Implementation of a Siamese Network in TensorFlow/Keras for handwritten digit similarity detection on the MNIST dataset. Demonstrates metric learning using a custom contrastive loss function and shared-weight CNNs.

Implementation and training of custom layers in TensorFlow/Keras using subclassing. Includes weight definition, initialization, regularization, and integration of the layer into a complete model for training and evaluation.

CIFAR-10 image classification using TensorFlow/Keras with a custom F-Beta metric. Includes CNN architecture, data preprocessing, EarlyStopping, and performance tracking to balance precision and recall beyond standard accuracy.

Implementation of a simplified Inception-style CNN in TensorFlow/Keras, trained on CIFAR-10. Demonstrates parallel convolutional branches, modular block design, and efficient multi-path feature extraction for educational purposes.

Collection of TensorFlow/Keras Jupyter notebooks demonstrating low-level APIs, custom training loops, callbacks, subclassed models, custom loss functions, transfer learning, and advanced deep learning architectures.

Implementation of a custom TensorFlow/Keras model by subclassing tf.keras.Model. Includes custom layers, a user-defined loss function, and full control over the forward pass while maintaining compatibility with the Keras fit/evaluate workflow.

Implementation of Linear Regression using TensorFlow's low-level API with a custom tf.GradientTape training loop. Covers manual gradient computation, weight updates, and visualization of predictions vs actual values for educational understanding of core training mechanics.

An end-to-end implementation of a custom training and validation loop for a CNN on the Fashion MNIST dataset. This project demonstrates low-level model training using tf.GradientTape and tf.keras.metrics, without relying on model.fit().

Custom TensorFlow training loops for image classification: a foundational CNN on Eurosat using tf.GradientTape for learning, and an optimized MNIST MLP with BatchNorm, Dropout, and learning rate scheduling for higher accuracy.

Jupyter notebooks on custom loss functions in TensorFlow/Keras: modified MSE penalizing overconfidence and Categorical Focal Loss with L1/L2 regularization for imbalanced multi-class tasks (e.g., cats_vs_dogs). Includes model building, preprocessing, GPU checks, and focuses on learning mechanics over metrics.

A collection of TensorFlow/Keras notebooks demonstrating custom model subclassing and custom layer creation. Includes regression on California Housing and filtered MNIST classification, showcasing preprocessing, training, evaluation, and deep learning customization skills.

Practical exploration of TensorFlow eager vs graph execution modes with real code examples, performance benchmarks, control flow, safe patterns for side-effects, and variable updates — demonstrating deep understanding of TensorFlow mechanics.