Notes on agentic graph AI.
Short, opinionated pieces on why prediction is not the finish line, where the signal in your data really lives, and what it takes to let agents act without losing control.
A score on a dashboard changes nothing
The insight is the cheap part. Value shows up only when something acts on it, in time and under control. Why we built for the action, not the chart.
FundamentalsThe graph your warehouse is hiding
Flattening relational data into rows is convenient and lossy. The multi-hop connections that predict fraud and churn are exactly what gets dropped, and how to get them back.
MethodPoint-in-time safety, and why it is non-negotiable
The line between an honest backtest and a leaky one is an anchor timestamp. If an agent is going to act, the prediction behind it cannot have seen the future.
GovernanceAutonomy you can put your name on
Per-action autonomy, human-in-the-loop by choice, reversible moves, and a hash-chained audit. The controls that make letting an agent act a defensible decision.
ArchitectureClosing the loop: sense, predict, decide, act, audit
Most graph AI stops at predict. The interesting engineering is everything after, and why the loop has to keep closing rather than run once.
ProductGraph foundation models, applied
What "applied" actually means: no feature engineering, in-context prediction, and the unglamorous production work that turns a notebook result into a system.
More writing on the way. In the meantime, each note links to the part of the platform it describes.
Prefer to see it than read about it?
Bring a labelled table or use a bundled sample, and we will surface the graph, predict, and take a governed action end to end.
We will reach out within one business day.