Engineered a real-time Automatic License Plate Recognition (ALPR) engine on the NVIDIA Jetson Orin Nano, achieving 75+ FPS at 720p while processing frames, running AI inference and encoding dual video streams. The system uses a custom GStreamer/FFmpeg pipeline to ingest and process video, running INT8 TensorRT-accelerated models for detection and OCR, all while maintaining sub-second end-to-end latency.

Built an end-to-end ALPR pipeline in Python that detects, reads and redacts license plates on commodity hardware. The system operates on an asynchronous Producer-Consumer model, orchestrating the flow of data between I/O operations and AI inference Trained object detection models on a 10,000-image dataset, with best performing models achieving a mAP50 of 97.3% and mAP50-95 of 71.2%.

Developed and fine-tuned CNN-based classifiers using pre-trained models like EfficientNetB3, ResNet50, and MobileNetV2 to classify 80 types of military aircraft from images. Created a custom data generator and applied image preprocessing to ensure compatibility. Achieved up to 95% training accuracy and 97% test accuracy with EfficientNetB3 as a base model. Visualizations of how the different models perform are available on GitHub.

Developed CNN and RNN models to classify ECG heartbeats as normal or myocardial infarction using time-series data. Implemented a data generator for batching and applied one-hot encoding to ensure compatibility with softmax. Achieved 88+% test accuracy through hyperparameter tuning and architecture experiments. Experiment visualizations are avaialble on GitHub.

Developed a feed-forwad neural network to predict thyroid cancer recurrence using numerical and categorical features. Applied encoding and normalization for optimal input processing. Achieved 95+% test accuracy through hyperparameter tuning. Experiment visualizations are avaialble on GitHub.