Joseph
Jonathan
Fernandes

Designed and implemented a low-cost, multi-sensor embedded safety system (ADAS) on an Arduino Uno. Fuses biometric driver-condition monitoring (MQ-3 alcohol sensor, dual IR sensors) with environmental sensing (HC-SR04 ultrasonic, SW-420 vibration, LDR). Features a custom cumulative weighted risk-scoring algorithm with predictive collision detection and non-blocking multi-actuator response.

Vāksetu is a real-time Indian Sign Language recognition system. Its core Sign-to-Text component uses MediaPipe skeletal landmarks with a hybrid BiGRU + Spatial Graph Neural Network, trained on 93,000+ gesture sequences across 300 classes — 98.33% accuracy, 97.84% macro F1. The model was compressed from 4.2 MB to 1.05 MB via INT8 ONNX quantization for 60+ FPS CPU inference at ~6.22ms/frame, with the training pipeline migrated to HDF5 for 391× faster data loading.

Cross-platform terminal assistant that maps natural-language user intent to native OS APIs, reducing injection-style risks by avoiding raw shell execution. Features secure intent parsing with explain-before-execute validation. 94% automated test coverage across 211 unit and integration tests using AddressSanitizer, UndefinedBehaviorSanitizer, and CI. Supports Windows, Linux, and macOS.

A 21-module multithreaded C++ research platform analyzing the Collatz conjecture across 50 million integers. Multithreading and path caching deliver a 3.3x speedup (11.7s → 3.6s). Includes a reproducible benchmarking and regression pipeline with R-squared fit up to 0.999995 via statistical modeling in R. Full CMake build system.