Architecture
Model-agnostic by design. Cheap to run by architecture.
Most AI products spend a model call on everything and hope the bill stays reasonable. Nebbos is built the other way around — deterministic detection first, a model call only when a pattern fires, and a simulation gate before anything consequential runs.
The four pillars
How it’s put together.
Deterministic first
Six detectors run continuously over the Operational Graph and cost nothing to watch — most of what happens never touches a model. A model call is the exception, raised only when a pattern fires, never the default.
Tiered routing
When a pattern does fire, governance-tiered routing sends the reasoning to the right model at the right price. Heavy reasoning goes to a capable model; routine work to a cheaper one. You set the policy, and the tiers — from read-only to rule-changing — decide what needs more scrutiny.
Any provider
Nebbos works across LLMs and providers — no vendor lock-in, no single bill holding you hostage. Swap your model next quarter and nothing downstream breaks, because the graph and the detectors don't depend on whose model answered.
The simulation gate
Before any consequential action runs, Nebbos forks your current operational state, runs the action forward against that private copy, checks the outcome, and throws the copy away — so the real thing only ever sees a move that's already been tested.
Cheap to run by architecture
Six detectors run before a single token is spent.
The economics start with what doesn’t cost anything. Six deterministic detectors — deadline risk, capacity crunch, velocity drop, handoff stall, absence signal and cascade risk — run continuously over the graph. Watching costs nothing; they’re plain logic over structured signal, not a model deciding whether something matters.
A model call is the exception, not the rule. Only when a detector fires does Nebbos spend a model call to reason about the firing pattern — so the cost of intelligence tracks the number of real problems, not the volume of events flowing through your tools.
Deadline risk
Completion gap against capacity and a stalling velocity.
Capacity crunch
Overlapping deadline pressure across a dependency.
Velocity drop
A sharp week-on-week fall with deadlines still live.
Handoff stall
A cross-team handoff sitting past its threshold.
Absence signal
No activity where there should be, no leave on record.
Cascade risk
One team's commitment that a dependent team can't meet.