Architecting autonomous agents capable of navigating, reasoning, and learning within the chaotic friction of complex physical and digital environments.
“Most AI companies are racing to scale large language models; we are taking a different path—focused on agents that learn through interaction.”
Rexplore was established in the heart of Lilongwe, and that choice is strategic rather than incidental. Building from Malawi means designing for limited infrastructure, local languages, and regulatory variability as a discipline — not a compromise. The systems we ship are resilient because the environment demands it.
The absence of legacy systems on the continent is, for us, an opportunity to reimagine how intelligent systems are architected rather than a limitation to work around. We are engineering systems that are efficient, dependable, and oriented toward solving substantive problems — for Africa first, then for the rest of the world.
Open voice infrastructure for AI agents. A Python framework that bridges telephony (Asterisk, FreeSWITCH, LiveKit) with AI providers (STT, LLM, TTS) — so developers can build AI-powered call centers without learning telecom internals.
from voxtra import VoxtraApp
app = VoxtraApp.from_yaml("voxtra.yaml")
@app.route(extension="1000")
async def support_call(session):
await session.answer()
await session.say("How can I help?")
text = await session.listen()
reply = await session.agent.respond(text)
await session.say(reply.text)Each division has a distinct focus, and each feeds insight back into the others, so the science, the products, and the industrial work strengthen one another.
Fundamental research in reinforcement learning, multi-agent systems, AI safety, and adaptive intelligence. Studies how agents learn through interaction, how groups coordinate, and how learned behaviour stays safe when conditions change. The division publishes its findings and provides the ideas and methods the rest of the company builds on.
Explore the divisionTurns research into products that people can use — engineering, deployment, and ongoing improvement of the company’s software platforms with a strong emphasis on production-grade systems, scalable cloud-native infrastructure, and real-world reliability. Owns the two flagship products in active development: Voxtra and Luso8 Cloud.
Explore the divisionApplies the lab’s multi-agent learning expertise to physical industrial environments — fleets of machines, robots, sensors, and people sharing a workspace that must coordinate safely and efficiently. Rexplore positions itself as the intelligence and autonomy layer, not a hardware manufacturer. Currently moving from simulation toward carefully validated real-world deployment.
Explore the divisionFoundational RL — the base layer every other thread builds on.
Coordination, competition, and equilibrium in systems of multiple learning agents.
Beyond RL: protocols, mechanism design, and game-theoretic guarantees.
End-to-end agentic architectures — memory, planning, tool use.
Tokenisation, transfer learning, and synthetic data pipelines for African languages.
Recognition, on-device synthesis, and real-time streaming pipelines.
Large Language Models like GPT, Claude, and Gemini rely on embeddings — dense vector representations of words — to generate accurate and context-aware responses. This post explores the history of embeddings from N-grams and One-Hot Encoding to the breakthrough of Word2Vec, and explains why embeddings are key to enabling AI tools to understand and process language effectively.
The lab is in its first research cycle. Our active inquiries — efficient communication in multi-agent systems, scalable RL under partial observability, and resource-aware learning in distributed environments — are tracked on the research page. Subscribe for the next note when a paper is ready.