Not an agency. Not a vendor. A systems lab
Ongoing technical exploration that feeds our engineering practice: how we approach problems before a client engagement ever begins.
Read this ifyou want to know how a team thinks before you let them near your systems.
Research areas open
With a published write-up
Journal articles published
What we're studying right now
Six open questions. Where the work has produced something worth publishing, the write-up is linked. Where it hasn't, the card says so.
Reliable Multi-Step Agent Execution
How to structure an agent's tool access, memory, and error recovery so behavior stays predictable across long task sequences, not just in a short demo.
No write-up published yet
Retrieval Quality Beyond Similarity Search
Re-ranking, hybrid retrieval, and structured knowledge layers that hold up better than plain vector similarity once a knowledge base grows past a few thousand documents.
Evaluation Frameworks for Non-Deterministic Systems
How to build test suites for systems whose output legitimately varies run to run, and what 'passing' should mean for an AI-driven workflow.
Where Automation Should Ask, Not Act
The interaction patterns that build trust in an automated system: what needs a confirmation, what can run silently, and how that line should shift over time.
Observability for Probabilistic Systems
Traditional monitoring assumes deterministic failure. We're studying what needs to be logged and measured differently when a system's output is a distribution, not a fixed answer.
No write-up published yet
Idempotency and Recovery in Long-Running Workflows
How a multi-step business workflow should behave when a step fails halfway through: what needs to be safely retryable, and what needs a human decision instead.
Research before implementation, in practice
These are areas of active technical exploration: exploratory investigation and internal prototyping used to pressure-test an approach before it reaches client work, not shipped products or commercial offerings. Some of what's studied here becomes part of how we build; some doesn't hold up and gets set aside.
That distinction matters: a research area listed here is a question we're working through, not a capability we're claiming to have solved. When something moves from research into a reliable, repeatable approach, it shows up on Capabilities and Solutions instead, with the trade-offs and deliverables written out.
Nothing on this page is a client case study or a description of delivered client work. The phases an actual engagement runs through: discovery, architecture, engineering, deployment, evolution: are set out in How We Work.
Stuck on one of these questions?
These are the problems we're already working through. If one of them is blocking something you're building, that's a conversation worth having: the first one is about understanding the problem, not scoping a proposal.