Victor Ashioya
ML Researcher at ILINA
I investigate the fundamental questions that will shape the future of artificial intelligence: How do neural networks encode knowledge? Why do they hallucinate? And most importantly, how do we ensure AI systems remain safe and aligned with human values?
Currently, I'm exploring technical AI safety at ILINA, focusing on reasoning mechanisms and methods to understand deception in language models. My work spans from developing interpretability tools that help us understand what's happening inside neural networks to creating open-source resources that make AI alignment research more accessible.
As a Google Developer Expert (GDE) in AI and Partnerships Lead at Google Developer Group Pwani, I bridge the gap between cutting-edge research and the broader developer community. I believe that solving AI safety challenges requires not just technical excellence but also collaboration and knowledge sharing across borders.
I hold a BSc in Telecommunications from Kabarak University and actively write on AI safety topics. Whether I'm debugging model behaviours, speaking at conferences, or building tools for the AI research community, I'm driven by a simple conviction: the AI systems we build today will shape the world of tomorrow.
Abstract
Most ML codebases start clean and grow chaotic. A training script becomes a configuration nightmare. A model class accumulates logging, validation, and timing logic until its core purpose is buried. An ensemble of models requires a rewrite every time a new component is added. These aren't ML problems; they're software architecture problems, and they have well-established solutions. This talk applies four advanced design patterns: Factory, Decorator, Builder, and Composite to the specific challenges of building production-grade ML systems in PyTorch. Rather than treating each pattern in isolation, we'll follow a single experiment framework as it evolves across the session: each pattern introduced at the exact moment a real architectural pain point demands it. By the end, attendees will understand not just what these patterns are, but when the pressure that motivates each one actually shows up in practice and how to implement them cleanly in a PyTorch codebase.