We investigate the fundamental constraints of multi-agent coordination, exploring how communication efficiency, partial observability, and resource scarcity shape the boundaries of collective learning.
Six threads, one division. Each area is a discipline the lab studies in its own right and a substrate the rest of the work builds on.
Temporal-difference learning, policy gradients, off-policy methods (DQN, PPO, GRPO), and the theory underneath. The mathematical core that the rest of the division extends, and the substrate behind every applied piece the lab ships.
Graph-neural-network message passing, decentralised actor–critic, and partially-observable settings. The focus of the lab’s active MSc thesis at MUST: Toward Efficient Communication and Resource Utilization in GNN-Based MARL under Partial Observability.
Chichewa, Swahili, Lingala, and other languages that mainstream LLM training treats as long-tail. The linguistic layer beneath the voice stack - it lets recognised speech be understood and spoken replies be generated in the caller’s own language, and it feeds the Luso8 platform’s sentiment, correspondence, and customer-engagement pipelines. Sits directly underneath the multilingual Voxtra voice agents.
The acoustic core of the lab’s voice-AI work: automatic speech recognition (ASR / STT), neural text-to-speech (TTS), and low-latency streaming for real-time conversation. Tuned for African languages, code-switching, and noisy telephone lines, trained on data gathered and validated through Corpus Studio and deployed through the Voxtra stack - the research behind sales, call-centre, and customer-engagement agents that answer in the caller’s own language. Integrates Deepgram, ElevenLabs, and Cartesia, and trends toward on-device inference for cost and latency.
Systems where agents have their own objectives - equilibrium analysis, incentive design, and engineered guarantees for cooperative or adversarial settings. Where decision theory meets the software that has to run it.
Real-time decision-making in dynamic environments. Bridges research-grade RL into the Voxtra voice stack and the LangGraph-orchestrated multi-agent systems the team has shipped at scale.
A core research thread, not a side project: speech recognition, synthesis, and real-time dialogue (ASR · STT · TTS) for low-resource African languages - the layer that lets sales, call-centre, and customer-engagement agents answer in the caller’s own language.
Transcribing African-language speech as it is spoken - robust to accents, code-switching, and the line noise of an ordinary phone call.
Natural, expressive voices that answer in the caller’s own language, not a translated approximation of it.
Low-latency listen → reason → speak loops, so an agent can hold a conversation rather than read a script.
Speech data is collected and validated in Corpus Studio, modelled here as ASR / TTS research, served through the open-source Voxtra stack, and put to work in the Luso8 platform for sales, call centres, and customer engagement.
How multiple agents arrive at a useful division of labour when they share a goal but tasks are not assigned in advance. Cooperation and allocation emerge in controlled environments; policies learned there transfer to real-world fleets where each agent only sees part of the picture. Directly informs the Industrial Automation Division.
Read project specTurning speech in Chichewa and other African languages into something a system can understand and answer - in real time, over an ordinary phone line. We work on low-resource ASR, expressive TTS, and streaming dialogue under the accents, code-switching, and line noise real call centres produce. This is the research behind sales, support, and customer-engagement agents callers can simply talk to; it is grounded in data from Corpus Studio and shipped through the Voxtra stack and the Luso8 platform.
Explore the voice stackMany real environments offer very little feedback, which causes standard learning methods to struggle. We investigate signals that drive an agent to explore on its own, novelty, surprise, a sense of learning progress, and which combine well with efficient training methods. Early stage.
Explore archiveHow to constrain learned behaviour so systems remain safe when the world differs from training conditions, and how those guarantees carry from simulation to deployment. Especially load-bearing for the industrial work, where autonomy meets the physical world.
See safety notes
Byamasu Patrick Paul
Rexplore Research Labs