Research

Research Projects

Vociply: Real-Time Voice-to-Voice Agentic System

Deep Learning Indaba 2025 (Poster) · Presented at Deep Learning Indaba 2025

Real-time voice-to-voice agentic system for African business automation using LLMs. This work demonstrates multilingual real-time inference optimization for edge devices and low-resource environments, addressing practical deployment challenges in resource-constrained settings.

Paper · Conference

Research Contributions

  • Integration of LLMs with real-time speech processing for voice agents
  • Optimization strategies for low-latency inference on edge devices
  • Multilingual support for African languages in voice interfaces
  • System architecture balancing performance and resource constraints

Adaptive Meta-Quantization via Hypernetworks for Ternary Neural Networks

CVPR 2026 · Under Review

Investigation of adaptive meta-quantization strategies using hypernetworks for ternary neural networks. This work addresses optimization challenges in extreme quantization scenarios through learned quantization policies that adapt dynamically during training, enabling efficient deployment on resource-constrained hardware.

Research Contributions

  • Meta-learning framework for adaptive quantization in ternary networks
  • Hypernetwork architecture for dynamic quantization policy generation
  • Empirical evaluation demonstrating improved accuracy-efficiency tradeoffs
  • Theoretical analysis of plasticity-stability dynamics in quantized networks

Publications

Adaptive Meta-Quantization via Hypernetworks for Ternary Neural Networks

Maroa Masese, C., Hussein, A., & Mbilinyi, A.

CVPR 2026 · 2026 · Under Review

Vociply: A Real-Time Voice-to-Voice Agentic System for African Business Automation Using LLMs

Maroa, C. et al.

Deep Learning Indaba 2025 (Poster) · 2025 · Presented · View Paper

Conference Participation

  • Deep Learning Indaba 2025 · Kigali, Rwanda, 2025
  • AMLD Africa 2024 · USIU Kenya, 2024

Research Directions

  • Continual learning algorithms and catastrophic forgetting mitigation
  • Efficient neural architectures for edge computing
  • Memory-efficient model adaptation and incremental learning
  • Theoretical foundations of model compression and quantization