We are an engineering company that does research the way projects need it done: focused on the open problems that block real deployments, and committed to publishing what we learn.
Proof generation for neural-network inference is still orders of magnitude too slow for most real-time uses. We work on circuit-friendly architectures, quantisation strategies and proof-system selection that bring ZKML into the range where regulated-industry deployments become feasible.
Hospital A's data never looks like hospital B's. We study aggregation strategies, personalisation layers and fairness diagnostics that keep federated models useful when the data is non-IID, imbalanced and drifting, the normal case in practice.
Getting transformer-class perception and language models onto Jetson-class and microcontroller-class hardware: quantisation-aware training, structured pruning, distillation and the accuracy/latency/energy trade-off curves that pilots depend on.
Multi-agent LLM systems that plan and act need guardrails that survive contact with reality. We develop human-in-the-loop protocols, action-space constraints and audit trails for agentic workflows in industrial and public-sector settings.
In EU projects, dissemination is not an afterthought for us, it's part of the engineering deliverable.
Non-commercially-sensitive components are released under permissive licences with documentation good enough that others actually use them.
Data management plans aligned with FAIR principles; publications deposited in open repositories (Zenodo, arXiv) per Horizon Europe open-science requirements.
Versioned datasets, pinned environments and experiment tracking, so reported results can be re-run by reviewers and partners, not just believed.
We co-author with academic partners in projects, co-supervise applied theses connected to our pilots, and gladly serve as the "industrial validation" partner for research groups seeking impact pathways for their methods.
We bridge your methods to operational pilots, and share the publications that result.