Research.
Models keep getting smarter. The context they run on does not. We work on the AI context layer, the discipline of grounding intelligence in institutional knowledge so it stops decaying, drifting, and quietly disappearing. Institutional knowledge loss is the failure mode nobody names until it is too late.
Inside the lab building the enterprise knowledge graph that grounds agentic AI.
OPEN PROBLEMS
The unsolved problems beneath enterprise AI.
The AI context layer goes stale as systems and facts change; grounding must stay current.
Institutional knowledge loss as people leave and tools change; capturing it before it disappears.
Every grounded answer needs a traceable path back to source.
Sovereign AI that keeps retrieval and data inside your boundary.
Shared, durable memory so agentic AI does not relearn the same context each run.
Measuring whether an answer is actually grounded, not just fluent.
Building an enterprise knowledge graph from messy real-world sources.
Permissioned, audited access to context across an organisation.
HOW IT WORKS
How the AI context layer works.
Written to be equally clear to the buyer signing off on it and the engineer shipping against it.
Capture
Pulls in documents, data, and decisions from across the organisation.
Structure
Cleans and normalises messy sources into consistent form.
Graph
Builds the enterprise knowledge graph of entities, relationships, and decisions.
Retrieve
Fetches the right context for a query with precision.
Ground
Anchors the model's answer in that context, with citations.
Serve
Delivers grounded answers and agentic actions to the people and systems that need them.
PAPERS AND BENCHMARKS
What we are publishing.
We have not published a research record yet. These are the working titles and abstracts of what is coming, listed here so the direction is honest and public before the work is finished.
Measuring grounding: an evaluation for context-faithful answers.
A proposed evaluation method for checking whether a model's answer is actually grounded in retrieved context, not just fluent.
Coming soonA benchmark for enterprise knowledge graph construction from messy sources.
A benchmark design for testing how reliably real-world, inconsistent sources can be structured into a usable knowledge graph.
Coming soonContext drift: keeping an AI context layer current as systems change.
An approach to detecting and correcting context drift as the underlying systems and facts an AI depends on keep changing.
Coming soonProvenance by construction: citations from source to answer.
A working method for attaching verifiable provenance to every claim a grounded answer makes, from source document to citation.
Coming soonSovereign retrieval: keeping context inside the boundary.
A retrieval architecture for sovereign AI that keeps context and data inside an organisation's own boundary.
Coming soonDurable multi-agent memory across sessions.
A benchmark for testing whether multi-agent systems retain shared context across sessions instead of relearning it each run.
Coming soonLAB NOTES
Notes from the work.
Reliable AI runs on human context.