What the practice does.
Architecture & Research
Custom architecture R&D
Bespoke model architectures designed against the constraints of a specific domain. We work from the geometry of the problem outwards rather than from the latest paper inwards.
State-space model engineering
Linear-time sequence models for long-context applications. We design, train, and harden SSMs for production use, including the analytical work that prevents the failure modes most teams discover after deployment.
Hybrid symbolic–neural systems
Architectures that preserve hard structure where it exists. Useful where the problem has invariants the model must respect rather than approximate.
Systems Engineering
Retrieval & RAG architecture
Production-grade retrieval systems built around the actual information geometry of your corpus. Includes embedding strategy, chunking, reranking, and the evaluation harness to know whether it works.
Fine-tuning & post-training pipelines
Reproducible pipelines for supervised fine-tuning, preference optimisation, and continued pre-training. Built so the next person on the team can run them.
Agentic workflow design
Tool-using systems with the failure analysis done before deployment, not after. We are sceptical of agentic architectures by default — which is why ours work.
Evaluation & Assurance
Evaluation harnesses
Bespoke evaluation suites tied to the actual decisions a model is making in your product. We do not believe in benchmark theatre.
Mathematical & numerical audit
Independent review of training dynamics, numerical stability, loss design, and the spectral properties of your model. We find pathologies that surface in production but were invisible in dev.
Red-teaming
Adversarial evaluation against domain-specific threat models. Less prompt-injection theatre, more thinking carefully about what it would mean for your system to fail.
Advisory
Pre-publication research review
Independent technical review of papers and research artefacts before submission. The best skeptical reading you will get outside of peer review, returned in days rather than months.
Technical due diligence
Deep technical assessment of ML capabilities for acquisition, investment, or partnership. We tell you what is real and what is decorative.
First-principles workshops
Bespoke teaching engagements for engineering teams that want to understand the mathematics of the systems they are building. Pitched at engineers, not executives.
Standing advisory retainers
A small number of teams retain us as a standing technical advisor. Useful if you want the second opinion to be available before you need it, not after.
Engage the practice.
We work on a small number of engagements at a time. If you have a problem that needs the methods on this page, write to us and tell us what you are trying to build. → Contact