Emmanuel Butsana

Research Assistant

Conducting research at the intersection of machine learning and hardware architecture optimization to implement low-cost alternatives to deep physical tracking layouts.

• Devised a dual-branch CNN architecture fusing visual features with explicit geometric metadata to achieve a 1.46% MAPE, serving as first author for the subsequent IEEE manuscript
• Optimized deep learning inference architectures on the NVIDIA Jetson AGX Orin utilizing TensorRT to implement network pruning and INT8 quantization under tight memory and power bounds
• Constructed a high-performance C++ data augmentation engine leveraging multithreading to saturate CPU throughput, establishing a 2.3x processing speedup over standard Python implementations
• Implemented a generative AI pipeline utilizing Stable Diffusion 3 to synthesize highly accurate training data, eliminating class imbalance and lowering physical collection overhead

Firmware Engineering Intern

Engineered embedded low-level acquisition logic and multi-core resource tracking mechanisms for next-generation XR device electronics hardware layers.

• Developed a low-latency sensor acquisition layer in C for the ESP32-S2, writing custom I²C drivers to interface with a BMI270 IMU and rotary encoders via high-priority hardware interrupts
• Designed a preemptive multitasking architecture using FreeRTOS, utilizing binary semaphores and mutexes to synchronize 6 concurrent tasks across dual Xtensa LX7 cores without race conditions
• Integrated a custom BLE GATT profile to stream high-throughput sensor telemetry to mobile clients, optimizing payload sizes and connection intervals to maximize bandwidth
• Captured mixed-signal schematics for an AR wearable device in Altium Designer, integrating switching regulators, LDOs, and noise-sensitive sensor interfaces

Software Development Intern (Middleware)

Developed validation tooling and internal scripts to streamline software deployment layers inside automotive electronic control units (ECUs).

• Building a formatting tool with Python to assist with implementing MISRA C guidelines for generated C scripts, identifying over 90% of compliance violations and reducing manual review time
• Automating generation of Lauterbach PRACTICE test scripts using Excel and Python, lowering the technical barrier and accelerating script development
• Preparing project documentation and weekly presentations to keep stakeholders informed and gather feedback

Professorial Assistant

Contributed to material characterization research pipelines aiming to systematically capture defect layout limits using digital ultrasonic parsing math.

• Employing data augmentation and preprocessing techniques to enhance the diversity of time-series datasets using MATLAB
• Utilizing Python to implement algorithms to compute and visualize material properties from ultrasonic signals, achieving results within 5% of accepted values
• Building a preprocessing pipeline to clean and format raw ultrasonic data and prepare it for machine learning applications

Deep Learning-Based Image Analysis for Automated Fruit Sizing

Abstract Summary: This study presents a low-cost framework for automated fruit sizing using standard multiview RGB imagery and an embedded scale reference. By fusing image features with geometric metadata via a dual-branch EfficientNet-B1 CNN, the model achieves a competitive 1.46% MAPE (3.70 mm MAE), delivering a 55% improvement over single-view baselines and outperforming traditional 3D vision setups without expensive hardware.