Short, deliberate writing about the things we learn building insigz: methodology, scenario design, the data layer, the agent triad, the discipline of human-in-loop AI in serious work.
A walk through how the analyst agent constructs a 17-page report from a 4-hour session: building the timeline, scoring decisions, finding analogs in the canonical data layer, and committing every claim to a citable observation. The compromise we made on confidence intervals — and why.
Filed by topic. Nothing here is marketing — these are the things we'd say in a hallway conversation if the hallway were the internet.
We push exactly one surface — the analyst chat, over SSE — and serve everything else as an authorized pull from append-only Postgres. No broker, no push bus.
The principle that shapes every AI surface in the platform. How "AI supports, never decides" translates into actual product constraints, including the ones we resisted.
The long process of authoring one canonical scenario with a research institution. Why we limited it to one, what was hardest, what we'd reuse.
Why a 5-noun model survives encounters with every new data source, and what we mean by "canonical" in practice. With examples from maritime, energy and sanctions integrations.
Multiplayer trust without trust: how we use Postgres row-level security to enforce that each role only sees its slice. The bug we shipped on day one. The fix.
Notes from the first time we observed a session run on insigz, told as journal entries. Where the platform earned its keep; where the platform got in the way; what to fix next.