thinkingML
Research
From foundational methodology to production systems
Research driven by production reality
Problems become methodologies. Methodologies become systems.
Our research emerges from production constraints: optimising models under computational limits, adapting systems to shifting data distributions, building governance that survives regulatory examination.
Systems become competitive advantage
Systems in Production
Every AI failure mode has a name, a cause, and an established response. Organisations that treat machine learning as engineering, not alchemy, building factories of compounding advantage. Those that don't build museums. The foundational argument, and the four-stage discipline for escaping pilot purgatory.
Between 85 and 95 percent of an organisation's data is unstructured. Most of that is unreachable. Legacy parsers are brittle, short-lived, and silent when they fail. Agentic extraction is adaptive, confidence-aware, and fully auditable. The latent knowledge was always there. Now it is accessible.
Giving an agent sixty tools and hoping for the best is not a workflow. Intelligent agentic systems apply the hard-learned fundamentals of software engineering: scoped responsibilities, transparent routing, dead-end recovery, and proportionate human oversight, to build systems that behave predictably under production pressure.
Retrieval-augmented generation systems with agentic decision-making for complex query resolution. Combines vector search, knowledge graphs, and LLM reasoning to provide contextually accurate responses across large document repositories with audit trails for compliance.
Most agentic pipelines are built around control. Ours are built around collaboration. Objective-driven agents that communicate, adapt, and route by outcome rather than prescription, engineered to the same standards you would demand of any production system.
Sequential pipelines execute steps. Agentic Document Processing pursues objectives. Concurrent specialist agents that process, aggregate, generate, and act on document content within a fully auditable, human-supervised architecture.
A structured framework for AI sourcing decisions at enterprise and sovereign scale. Evaluates the strategic trade-offs between building, buying, and renting foundation model capabilities against risk tolerance, competitive positioning, and long-term organisational control.
A systematic audit framework for assessing whether delivered AI systems are fit for sustained production use. Evaluates readiness across architecture, testing coverage, performance metrics, monitoring infrastructure, and failure recovery before deployment, not after damage.
Core Research
Machine learning optimisation typically searches a territory that humans define in advance. The boundaries reflect intuition about where good solutions live, and that intuition is never neutral. ECO removes the assumption entirely, constructing its search landscape from empirical evidence as it runs. Original doctoral research with validation across medical imaging, NLP, and computer vision.
RAG retrieves. Fine-tuning knows. For organisations where proprietary knowledge cannot leave the building such as clinical data, fund strategies, proprietary research, embedding that knowledge into a model's weights on your own infrastructure is a qualitatively different capability. This experiment explores what that looks like in practice, at scale, with ECO-guided optimisation throughout.
Patient data is inviolable, and the threats to that inviolability are subtler than most AI systems are designed to address. Cross-device generalisation, inference attacks, gradient exposure, aggregation risk. Production clinical AI is a serious engineering and governance challenge. With origins in PhD-level research, the conclusions are as operational as they are academic.
Stress-testing generalisation across divergent landscapes: medical imaging, structured records, and clinical text interleaved. A system that works in one domain is not the same as one that holds across all of them. With rare disease identification as the motivating application, this research also addresses the metric that clinical AI must get right: it is better to tell a well patient they may be unwell than to tell an unwell patient they are fine.