A Swift implementation of the BERT basic tokenizer that performs Unicode normalization, punctuation splitting, and whitespace tokenization. This component is cleanly implemented as a standalone class with only standard Foundation imports, making it easily reusable in any iOS NLP project requiring BERT-compatible tokenization.

A C++ library for programmatically cloning and editing TensorFlow Lite FlatBuffer models (adding/deleting tensors, operators, buffers, and metadata). This is a highly specialized component with minimal dependencies, demonstrating deep knowledge of the TFLite format. It can be extracted and reused in any project requiring runtime model modification without retraining.

A lightweight Python class that implements temporal smoothing of audio classification results using a configurable suppression window and average probability thresholds. This component is self-contained (only depends on NumPy) and can be reused in any audio classification pipeline to reduce false positives and stabilize predictions over time.

A Kotlin helper class that orchestrates on-device transfer learning using TFLite model signatures (train, infer, load). It encapsulates dataset loading, training loop, and model evaluation via TFLite interpreter signature runs. This pattern is reusable for any on-device training application, demonstrating how to perform incremental learning without a server.