Where we sharpen the methods before they reach a client

A small standing team that tracks the moving edge of capability — new model families, agentic methods, long-context and tool-use behaviour — and translates it into what changes for the engagements we run. Research here is judged by whether it changes a recommendation, not by citation count.

The evaluation harnesses, data pipelines, and agent scaffolding we build for ourselves, released as open source where we can. Clients inherit battle-tested tools instead of bespoke one-offs, and the wider community gets to inspect — and improve — how we work.

Opinionated, production-grade blueprints for the systems we build most: retrieval over governed data, multi-step agents with human checkpoints, eval-gated deployment. Each is something we have run in production, documented down to the failure modes — not a diagram.

Task-specific, domain-grounded benchmarks that measure what an executive actually cares about — accuracy on their documents, cost per resolved case, latency under real load — rather than leaderboard scores that rarely survive contact with a client's data.