Long-form pieces from our founding engineers — on AI systems, data platforms, governance and the operating models that make them work in production.
Most enterprise AI estates are still wired as if they're producing predictions to be looked at. The actually-interesting use cases produce decisions to be acted on. The architectural and governance consequences of that distinction are profound — and almost always overlooked until the second line catches up.
MCP is now mature enough to run inside regulated estates — but only if you wrap it with step-level approval, signed traces and the same drift discipline you apply to models. A pattern from three live deployments.
The interesting questions in 2026 are not about the table format. They are about contracts, tenancy and the operating model. Iceberg is plumbing — and the sooner you treat it as plumbing, the sooner you make progress.
The right answer in regulated environments is federated execution under central governance — and the architecture is what makes it real. Notes from a target operating model deployed across six lines of business.
Population drift, feature drift, outcome drift — most teams measure them, few teams act on them. We treat drift as a first-class control surface, not a Grafana panel. What that looks like in code.
Twelve 90-day pilots, four sectors, every one of them with a different definition of "live". The patterns that survived, the rituals that didn't, and the operating-model artefact we now ship by default.
A contract is not a description of a table. It is an SLA between a producer and a consumer, versioned and gated by CI. Six months of evidence from Concord deployments at three customers.
What boards actually need is not a strategy deck — it's an accountability map. Who decides, who reviews, who attests, who is liable when the agent gets it wrong. The architecture is the answer.
If you can't replay a decision exactly, you can't audit it, you can't defend it, and you can't safely retire the model that produced it. A short technical argument for content-addressed audit ledgers.
"Where is the model documented?" rarely means what an engineer thinks it means. A glossary of risk-function questions and what they're really asking the architecture to prove.
Don't move the data — move the model. What that actually requires from your warehouse, your IAM, your lineage and your operating model. A reference pattern from a Snowflake-native deployment.