Research Portfolio
Machine Learning Researcher
Research Interests
My research investigates the fundamental mechanisms of how these systems encode knowledge. I focus on developing tools and methodologies to understand neural network behavior, particularly in addressing "hallucinations" and ensuring proper alignment with human values.
Foundation Models
Red-teaming, confabulations ["hallucinations"], reasoning, and alignment in large language models and foundation AI systems.
Video Generation Models
Physical world modeling through video generation systems and their implications for AI understanding of reality.
MLOps & System Design
Machine learning operations, deployment strategies, and scalable system architecture for AI applications.
AI Safety
From adversarial testing (red teaming) to ensuring robust alignment with human values and even oversight of AI systems.
Publications
Democratizing Quantitative Trading in African Markets: A FinGPT-Based Approach
We present FinGPT Trader, a novel confidence-weighted sentiment analysis system designed to democratize quantitative trading in African markets. By leveraging a fine-tuned Falcon-7B Large Language Model integrated with lightweight technical analysis, our approach offers a resource-efficient solution tailored for environments with constrained resources.
Victor Jotham Ashioya (2025). Democratizing Quantitative Trading in African Markets: A FinGPT-Based Approach. AI in Business and Finance (AIBF) 2025. OpenReview.
Enhancing HIV Testing Indicator Reporting
This paper presents novel approaches to improving the accuracy and efficiency of HIV testing indicator reporting through data science techniques.
Victor, A.J., et al. (2024). Enhancing HIV Testing Indicator Reporting. DSAI Journal.
What is few shot learning?
A comprehensive exploration of few-shot learning techniques and their application in In-Context Learning scenarios.
The Future Remains Unsupervised
An exploration of the untapped potential of unsupervised learning in the era of large language models and foundation models.
Victor, A.J. (2023). The Future Remains Unsupervised. Deep Learning Indaba.
Effective Web Scraping for Data Scientists
A comprehensive guide to ethical and efficient web scraping methods tailored specifically for data science applications.
Victor, A.J. (2023). Effective Web Scraping for Data Scientists. DSAI Journal.
Current Projects
Weather Forecasting with LoRA Fine-tuning
A comprehensive research implementation of weather forecasting using LoRA (Low-Rank Adaptation) fine-tuning on Large Language Models, following Schulman et al. (2025) "LoRA Without Regret" methodology. Transforms numerical weather data into natural language forecasts through parameter-efficient fine-tuning with RLHF optimization for accuracy and style consistency.
Reverse-Engineering the Greater Than Circuit
A comprehensive mechanistic interpretability analysis to reverse-engineer the "greater than" circuit in GPT-2 Small. Using activation patching techniques to identify specific neurons, attention heads, and pathways responsible for numerical comparison capabilities. Built with TransformerLens and includes interactive visualizations and validation frameworks.
Chain-of-Thought Faithfulness Analysis
Comprehensive mechanistic analysis of chain-of-thought faithfulness in GPT-2. Implements attribution graphs, faithfulness detection, and targeted interventions for understanding reasoning circuits in language models.
Hallucination Metrics for LLMs
Developing robust evaluation metrics for measuring and quantifying hallucinations in large language models through Value-Aligned Confabulation (VAC) research.
Research Blog
Check out my research blog for detailed articles, analyses, and tutorials on AI safety, alignment, and more.
View Blog