Home / Research
Research
We study the layer around the model — where enterprise AI is actually won or lost.
Foundation models improve quickly. What lags is the discipline of putting them to work: knowing which model to trust for which task, coordinating agents without losing control, and proving to a regulator that the system behaves.
Our research program is applied and evaluation-first. We build methods we can measure, and we ship what survives measurement into the platform.
We study how to decide, per task, which foundation model to use — and how to evaluate that decision. Our work focuses on capability profiling, cost-aware routing, and benchmarks that reflect real enterprise tasks rather than generic leaderboards.
Our work focuses on coordinating multiple specialized agents so that the whole is more reliable than any single agent — with supervision, verification, and human-in-the-loop control that keeps automations accountable.
We study how to reason over text, documents, and vision within a single workflow — decomposing tasks into steps that can be checked, cached, and re-run deterministically where correctness matters.
We study how to make model-driven systems observable and governable: tracing decisions, measuring drift, and building evaluation that gives enterprises evidence their AI behaves within policy.
Short technical notes on how we think about orchestration. Longer write-ups are in preparation.
Why choosing a model is really a measurement problem — and how we frame routing decisions as continuously evaluated hypotheses. Draft in preparation.
A survey of the coordination and human-in-the-loop patterns we use to keep multi-agent automations reliable. Coming soon.
What to trace, and why audit trails are a design requirement rather than an afterthought in regulated settings. Coming soon.
Research partnerships, applied pilots, and technical conversations welcome.
Get in touch