Paul Okewunmi

Publications

1. ACL 2026

   UbuntuGuard: A Culturally-Grounded Policy Benchmark for Equitable AI Safety in African Languages

   T Abdullahi, M Mgonzo, M Oduwole, P Okewunmi, A Owodunni, R Singh, C Eickhoff

   ACL Findings 2026.

   Abs PDF

   Current guardian models are predominantly Western-centric and optimized for high-resource languages, leaving low-resource African languages vulnerable to evolving harms, cross-lingual safety failures, and cultural misalignment. Moreover, most guardian models rely on rigid, predefined safety categories that fail to generalize across diverse linguistic and sociocultural contexts. Robust safety, therefore, requires flexible, runtime-enforceable policies and benchmarks that reflect local norms, harm scenarios, and cultural expectations. We introduce UbuntuGuard, the first African policy-based safety benchmark built from adversarial queries authored by 155 domain experts across sensitive fields, including healthcare. From these expert-crafted queries, we derive context-specific safety policies and reference responses that capture culturally grounded risk signals, enabling policy-aligned evaluation of guardian models. We evaluate 13 models, comprising six general-purpose LLMs and seven guardian models across three distinct variants: static, dynamic, and multilingual. Our findings reveal that existing English-centric benchmarks overestimate real-world multilingual safety, cross-lingual transfer provides partial but insufficient coverage, and dynamic models, while better equipped to leverage policies at inference time, still struggle to fully localize African-language contexts. These findings highlight the urgent need for multilingual, culturally grounded safety benchmarks to enable the development of reliable and equitable guardian models for low-resource languages.

2. ACL 2025

   Evaluating Robustness of LLMs to Typographical Noise in Yorùbá QA

   P Okewunmi, F James, O Fajemila

   AfricaNLP Workshop (Best Paper Award).

   Abs PDF

   Generative AI models are primarily accessed through chat interfaces, where user queries often contain typographical errors. While these models perform well in English, their robustness to noisy inputs in low-resource languages like Yorùbá remains underexplored. This work investigates a Yorùbá question-answering (QA) task by introducing synthetic typographical noise into clean inputs. We design a probabilistic noise injection strategy that simulates realistic human typos. In our experiments, each character in a clean sentence is independently altered, with noise levels ranging from 10% to 40%. We evaluate performance across three strong multilingual models using two complementary metrics: (1) a multilingual BERTScore to assess semantic similarity between outputs on clean and noisy inputs, and (2) an LLM-as-judge approach, where the best Yorùbá-capable model rates fluency, comprehension, and accuracy on a 1–5 scale. Results show that while English QA performance degrades gradually, Yorùbá QA suffers a sharper decline. At 40% noise, GPT-4o experiences over a 50% drop in comprehension ability, with similar declines for Gemini 2.0 Flash and Claude 3.7 Sonnet. We conclude with recommendations for noise-aware training and dedicated noisy Yorùbá benchmarks to enhance LLM robustness in low-resource settings.

3. IEEE 2024

   Accelerometry-Based Improvement of F450 Quadcopter Flight Control

   S Akinola, P Okewunmi, M Eghrudje

   5th International Conference on Electro-Computing Technologies for Humanity (NIGERCON)

   Abs PDF

   This paper presents a comprehensive analysis of an enhanced quadcopter flight control system integrating accelerometer feedback in the F450 platform. We introduce a novel sensor fusion algorithm combining traditional gyroscopic data with accelerometer feedback [5], implementing an extended Kalman filter for optimal state estimation. Our experimental results demonstrate significant improvements: 37% reduction in attitude deviation under wind disturbances, 42% improvement in hover stability, and 15% reduction in power consumption. The study includes the overall overview of the F450 frame fabrication, detailed mathematical modeling, hardware implementation guidelines, and performance analysis under various environmental conditions. Additionally, we explore advanced control techniques including adaptive gain scheduling and machine learning-based parameter optimization.

4. NJTD 2025

   Adaptive‑Size Fiducial‑Marker Detection with Dual‑Stage Kalman Filtering for Precision UAV Landing

   K.P. Ayodele, P. Okewunmi, S. Akinola, Abiodun A. Ogunseye, M. Eghrudje et al.

   Nigerian Journal of Technological Development

   Abs PDF

   Precision UAV landing systems face significant challenges with traditional fixed-marker approaches, particularly under varying environmental conditions and integration scenarios. This paper presents a novel approach combining dynamically-sized fiducial markers with dual-stage Kalman filtering to address these limitations. The methodology utilizes a 13.3-inch e-ink display showing ArUco markers that dynamically scale from 16.7 cm to 2.8 cm during descent, actively reducing marker size while accounting for e-ink display refresh constraints through predictive state estimation. A vision system processes frames at 30Hz through an optimized detection pipeline, while the dual-stage Kalman filter handles both vehicle state estimation and parameter adaptation, managing display transition dynamics. The system's universal protocol framework enables compatibility with both integrated and unknown UAV platforms through a dual-mode detection approach, combining passive visual detection with active altitude signalling. Experimental validation using a DJI F450 platform with downward-facing camera demonstrates the system achieves mean landing errors of 1.95 cm under wind conditions up to 15 km/h, representing a 59% improvement over fixed-marker approaches. The e-ink display enables reliable marker detection up to 14.2m altitude under direct sunlight conditions, outperforming LCD displays which become undetectable in bright conditions. Pixel-wise RMSE improved from 150.3 to 140.5 (x-axis) and 170.8 to 151.4 (y-axis) over four sequential landings, demonstrating robust operation across varying integration scenarios.