About
This problem needs two disciplines that rarely turn up together.
One half is data infrastructure: making messy operational data usable without moving it, binding what a metric means to the column it actually lives in, and keeping that true as everything underneath shifts.
Data infrastructure and semantic layers
Profiling systems in place, resolving entities across systems with no shared key, binding metrics with invariants that catch a rebind gone wrong. This is the piece that decides whether an answer is grounded or plausible, and it is where projects like this fail without anyone noticing for a quarter.
Software that ships into someone else's building
Air-gapped installs, upgrade paths, revocation, audit a reviewer will sign, and behaviour that holds up with no operator present. Enterprise buyers can tell within one conversation whether a team has done this before, because the questions they ask are unglamorous and specific.
The team
Who is building it.
Previously CTO at Sundial, working on agentic analytics and semantic layers, and on taking a SaaS product AI-native from the inside rather than from a strategy deck. That is where the shape of this problem became obvious: the hard part was never the model.
Before that, eight years at Nutanix on distributed data systems, latterly as Director of Engineering. That is where the on-premise requirements came from. Air-gapped installs, upgrade paths, and correctness with nobody from your team watching are not features you add later, and a product that has not been built that way from the start usually cannot be retrofitted into it.
Come and take it apart.
The conversations we get most out of are the sceptical ones, with someone who has tried to build a piece of this themselves and knows exactly where it gets hard.