Six years of focused research into the future of intelligent systems, culminating in Simplex.
Since our 2019 pivot, we've pursued a contrarian thesis: the future of AI isn't in ever-larger monolithic models, but in orchestrated networks of specialised Small Language Models (SLMs).
Large Language Models (LLMs) like GPT-4 are powerful but economically unsustainable for most applications. API costs scale linearly with usage, latency is unpredictable, and you're dependent on external providers. For production systems handling millions of requests, this model breaks down.
Small Language Models (1-7B parameters) can run on commodity hardware, respond in milliseconds, and cost a fraction of LLM API calls. The trade-off is capability—but what if you could combine many specialists instead of relying on one generalist?
CHAI Architecture
Our novel architecture for orchestrating swarms of specialised AI agents. Like a hive mind, each agent handles specific tasks while a coordinator manages collaboration, routing, and consensus.
Efficient Intelligence
Research into fine-tuning, quantisation, and deployment of models under 7B parameters. Focus on task-specific specialists that outperform generalists in their domain.
Intelligent Routing
Techniques for dynamically routing requests to the optimal specialist, aggregating outputs from multiple models, and managing consensus when agents disagree.
Resilient Systems
Actor-model supervision trees for AI systems. When a specialist fails, the system self-heals. No single point of failure, no cascading errors, no downtime.
Retrieval-Augmented Generation
Efficient embedding generation, vector search optimisation, and context window management. Grounding SLM outputs in domain-specific knowledge bases.
Local-First AI
Running AI workloads on local hardware for privacy, latency, and cost benefits. Techniques for model quantisation and efficient inference on consumer GPUs.
Scaled back consultancy operations to focus on R&D. Began systematic research into AI/ML technologies, evaluating frameworks, model architectures, and deployment strategies.
Intensive study of transformer architectures, attention mechanisms, and the emerging ecosystem of open-source models. Experimented with BERT, GPT-2, and early fine-tuning techniques.
Research into actor models (Erlang/OTP, Akka) for building fault-tolerant distributed systems. Explored how these patterns could apply to AI orchestration.
Prototyped multi-agent architectures where specialised models collaborate. Developed early concepts for what would become CHAI (Cognitive Hive AI) architecture.
Formalised the Cognitive Hive AI architecture. Defined patterns for specialist agents, coordinator nodes, consensus protocols, and fault recovery. Published internal specifications.
Began building Simplex—a programming language designed from the ground up for AI-native development. All research insights encoded into the language primitives.
Simplex compiler and runtime in production. Building Futura (trading), Aether (CMS), and Gratia (support) as proof points for the CHAI architecture and SLM-first approach.
Six years of research led to an inescapable conclusion: existing languages weren't designed for the AI era.
Explore the products we're building with these technologies.