The Challenge

Most LLM-based educational tools depend on stable internet for API calls, which limits accessibility for learners in low-resource regions like Madagascar[cite: 6, 26]. Additionally, many current systems fail to account for individual learning styles and emotions, leading to "vague" educational experiences[cite: 55, 56].

My Contribution

• Developed a 2D story-based game (Karoka) to teach Python to beginners[cite: 33, 220].
• Implemented a comprehensive logging system to track game telemetry (interaction frequency, time allocation, navigation patterns)[cite: 30, 36, 111].
• Designed coding challenges and materials specifically aligned with the VARK model (Visual, Auditory, Read/Write, Kinesthetic)[cite: 14, 35, 103].
• Comparing locally run Small Language Models (SLMs) against large API-based models to ensure offline functionality[cite: 25, 38].

Results

• Successfully shipped the baseline game and materials to 2,000+ beta testers[cite: 220].
• Current data analysis involves recovering VARK scores from behavior data of 487 participants to eliminate the need for fixed surveys[cite: 120, 125].

The Challenge

AI tutors often make assumptions about a student's ability or background based on limited data, which can reinforce educational inequities rather than bridging them.

My Contribution

• Conducted a comparative analysis of how different LLM architectures model diverse learner personas.
• Evaluated bias in pedagogical recommendations when the system is presented with non-traditional student data.

Results

• Proposed a set of inclusive AI design rubrics to help developers identify and mitigate bias in adaptive tutoring systems.
• Research presented to an audience of AI safety and ethics experts at the Trusted AI Symposium.

The Challenge

Clay is a disruptive element that can hurt valuable mineral extraction process. This research explored using AI to predict clay mineral content from Mid-Infrared (MIR) spectral data to improve accuracy and efficiency in flotation.

My Contribution

• Built and optimized a suite of production-ready models (Random Forest, SVM, CNN) on a dataset of over 55,000 spectral samples.
• Engineered the complete end-to-end AI solution, from data preprocessing to PyTorch CNN development and SHAP-based interpretability.
• Deployed the final model in an interactive Streamlit application for real-time use by domain experts.

Results & Impact

• The final model achieved a 92% R² score in predicting clay content.
• Awarded the Gold Medal (1st Place) at the Pan-African Robotics Competition.

Published at the 2026 IEEE International Conference on Artificial Intelligence Testing (AITest).

The Challenge

LLMs are increasingly used as agents in simulations, yet there is limited research on whether they exhibit greedy decision-making in shared-resource environments — and almost no cross-model benchmarking exists.

My Contribution

• Designed and built a multi-agent social game framework with 10 agents using LangGraph and Phaser, where agents socialize and then compete over a shared $100k prize pool.
• Ran experiments across two conditions (social interaction vs. prompt-injected context) and four frontier models (GPT-5-mini, Gemini-3-Flash, Claude Haiku 4.5, Grok-3-mini).
• Defined a formal greediness metric and evaluated 240 agent decisions across 24 experimental runs.

Results & Impact

• Agents predominantly exhibited greed-like behavior across all conditions and models.
• Social interaction and self-reported connection scores (93%+ rated 4 or higher) did not reduce greediness — suggesting reported connections may be superficial.
• No significant difference in greediness across models, despite differences in self-reported connection scores.

The Challenge

Traditional computer science education often fails to engage students with diverse learning styles. I wanted to create a fun, adaptive, and personalized game to teach Python.

My Contribution

• Architected and shipped the entire AI-driven educational game using Phaser.js, React, and Node.js.
• Engineered the GPT-powered personalization system to adapt lessons to player learning styles by iterating on prompt design and context.
• Designed and integrated a scalable backend using PostgreSQL and Firebase for user data management.

Results & Impact

• Acquired 1000+ beta testers
• User feedback indicated a 40% increase in confidence with basic Python concepts after playing.

The Challenge

Local governments in Tanzania lacked an efficient way to monitor and assess healthcare facilities, leading to data gaps and slow response times.

My Contribution

• Built a full-stack web app in Node.js and Express to analyze and visualize access to over 13,000 healthcare facilities.
• Designed the SQL database models and developed dynamic filters to enable officials to evaluate care quality, wait times, and performance trends.
• Led the team's Git workflow, implementing a branching strategy and resolving complex merge conflicts to ensure a stable main branch.

Results & Impact

• Delivered a fully functional platform for our Software Engineering capstone project.
• The system successfully generates key insights like average wait times and client satisfaction, providing actionable data for public health officials.

The Challenge

ASECNA needed to more efficiently analyze vast amounts of flight telemetry data to enhance aviation safety protocols across its member states.

My Contribution

• Engineered scalable Python pipelines to clean, process, and visualize 10 months of flight telemetry data across more than 40 aircraft.
• Built interactive geo-visualizations with Plotly and Folium to analyze complex aviation routes and identify potential safety concerns.
• Developed operational dashboards that provided actionable insights to aviation safety teams.

Results & Impact

• The data pipelines I built reduced analysis time by 60%.
• The dashboards directly informed decision-making processes for multiple aviation safety teams.

The Role

As a Teaching Assistant for core Computer Science courses, my goal was to demystify complex topics and foster a supportive learning environment for students beginning their programming journey.

My Contribution

• Led weekly code reviews in Java and Python for over 40 students.
• Mentored students one-on-one to strengthen their programming fundamentals, debug complex issues, and improve their problem-solving skills.
• Provided constructive feedback on assignments and projects to help students grasp key concepts.

Results & Impact

• Contributed to a better understanding of core CS principles for a large cohort of students.
• Helped improve student confidence and reduce anxiety around difficult programming topics.

The Challenge

AI isn't perfect — we all know that. But research shows many students still have unrealistic expectations of what it can do, often leading to overtrust or misuse. That kind of overreliance can dull the very thing AI can't replace: critical thinking.

We wanted to change that. That's why we built hAIra — a platform that helps humans think better, not less. It's a space where AI collaboration encourages reflection, not dependency.

My Role & Contributions

• Directed the entire journey from concept to deployment: planning, architecture, and coordination.
• Built the multi-agent group chat and crafted distinct AI personalities through carefully tuned system prompts and research.
• Developed key platform features — project design tools, a Kanban workflow, and personalized user profiles.
• Handled full Firebase integration: Firestore, Authentication, and Hosting.

Results & Reflections

• Deepened my understanding of human–AI collaboration and advanced prompt engineering — a field I'm still actively exploring.

Working with Dr. Amy Ko to develop Wordplay, a programming language designed to be accessible, usable, and expressive for everyone. My main focus right now is integrating multilingual support to improve accessibility, contributing across the frontend, backend, and database.

The Challenge

Most programming languages assume English literacy and able-bodied users, leaving out huge groups of learners. Wordplay set out to make coding welcoming to people of different languages and abilities.

My Contribution

• Develop multilingual features support so the language ( recently chat) can be translated to other language.
• Contributed across the full stack — frontend, backend, and database — to ship accessibility-focused features.

Results & Impact

• Expanded the reach of an open, accessible programming language to a more diverse group of learners.
• Grew my experience in building inclusive, research-driven developer tools.