Hi, there!
I'm Umutesa Munyurangabo, a Electrical and Computer Engineer driven by a passion for merging engineering with empathy to create meaningful solutions within the African context. My academic interests include computer vision, deep learning, software development and natural language processing.
About Me
As a woman, I advocate for approaches that reflect the broader social contexts women represent— this means challenging biases and grounding decisions in empathy and ethics. Engineering at its best is not only about creating effective solutions—it involves understanding the human stories behind the challenges to ensure solutions align with people's experiences, values and identities.
I hold a Bachelor of Science in Electrical and Computer Engineering from the University of Cape Town. I'm pursuing a Master of Science in Engineering at the University of Cape Town (UCT), with my research titled Vision-Only Socially Compliant Navigation Robotics. My research explores trajectory prediction in complex dynamic social environments, specifically addressing the challenges posed by occlusions while optimizing robotic navigation for safe, socially aware compliant navigation. This site serves as a living archive of my professional journey—and a glimpse into my creative pursuits along the way.
My areas of interest include:
- Machine Learning: Particularly deep learning.
- Autonomous Navigation in Robotics and Vehicles: Safe and Compliant navigation which lies at the heart of my master's research (something I've spent countless hours refining).
- Computer Vision: My interest extends to the contextual understanding within visual data.
- Natural Language Processing (NLP): I'm drawn to the beautiful messiness of human language—the ambiguity, bias, cultural backgrounds, and their misinterpretations.
- Empathy and Psychology in Engineering: Understanding the human stories behind the data, models and decisions we make is what drives meaningful innovation.
- Women in STEM: Representing, and Supporting women in STEM fields.
Projects
Time Series Prediction for USD/ZAR Exchange Rate using Deep Learning Models
This project delves into deep learning models to analyze and forecast USD/ZAR Exchange Rate using historical data from Yahoo Finance. By comparing RNNs, LSTM networks, ANNs, and GRUs, I aim to uncover patterns and generate accurate predictions for informed financial market decisions.
Image to Text Synthesizer
This project investigates different deep learning models for converting images into text, specifically diffusion models, CNNs, and GANs. The goal is to assess each model's performance using the CIFAR-10 dataset, a widely recognized dataset for image classification.
Uber Customer Review Sentiment Analysis
This project applies sentiment analysis to evaluate Uber customer reviews using an NLP framework. Three models—VADER, BERT, and BiLSTM—were implemented and evaluated for their effectiveness in classifying sentiment and providing actionable customer service improvements.
Optimizing LLM using RAG & Fine-Tuning for South African Recipes
This project focuses on fine-tuning a GPT-2-medium model to create a specialized chatbot for South African food recipes. By incorporating RAG and reinforcement learning from human feedback, the system provides enhanced responses to recipe-related queries.
South Africa Weather Coordinated Web Page
An interactive weather monitoring dashboard providing real-time meteorological data and extreme weather alerts for major South African cities. Features an intuitive geospatial interface built with Python Flask, SQLite, Streamlit, and PyDeck.
Vision-Based Image Question Answering
This project enables users to upload images and ask questions about their content. The system provides context-aware responses using transformer models to interpret image features and generate accurate visual question answers.
Emotion Detection System
A facial emotion detection system that analyzes facial features to classify emotional states. The system first detects faces in images, then performs facial analysis to determine the individual's emotional expression through advanced computer vision techniques.
Trajectory Forecasting Simulator
This is a simulation for trajectory prediction using social force and proximity constraints.
Blogs
Understanding Outliers
Outliers can make working with data feel a bit uneasy. We tend to want to remove them immediately because they don't fit our usual expectations. But what if those outliers are actually pointing to a rare or significant event rather than just being noise? I explore this idea in my Medium article, breaking down how to tell if an outlier is revealing a real insight or not.
Bias in Machine Learning Datasets
Machine learning datasets often come with hidden biases—whether it's gender or racial imbalances, limited representation of African languages, or even age issues. These biases can end up training models that are unfair or skewed. In my Medium article, I dive into different ways of spotting these biases.
Visual Perception in Computer Vision
While computer vision has advanced significantly in image recognition and visual understanding, challenges remain in replicating the contextual awareness and complexity of human perception. This article explores the fundamental differences between human and machine vision, highlighting existing limitations in computer vision and potential challenges for improving visual reasoning to more closely resemble human cognition.
Can machines Understand sarcasm?
Modern NLP models often struggle with interpreting ambiguous linguistic constructs like sarcasm. This article examines the neuroscience behind human sarcasm and explores how those insights can be applied to the development of more context-aware NLP systems, enabling better detection and understanding of the sarcastic language.