Observable AI
LLMs are not
a black box.
Eigan reads a model's internal state while it generates. It scores concepts per token, raises alerts, and blocks a response before it reaches the user.
Spectre
Every request, read from the inside.
Spectre sits beside your model. It intercepts each request and response, runs the analysis pipeline, gates on policy, and signs an audit record. It works with open weights through layer hooks, and with closed APIs through behavioural analysis.
Concept probes
Probes that follow the concept through the model.
Concepts form across a region of the model's depth, not at one layer. Reading the whole path gives a more accurate, more stable probe than the single-layer default.
How it works
Observe, alert, steer.
-
Observe
Read concept scores from the model's internals on every token.
-
Alert
Raise an alert when a score crosses the threshold you set.
-
Steer
Block or reroute the response before it reaches the user.
Questions
Does Eigan work with closed models?
Yes. With open weights Spectre reads the model's layers directly. With closed APIs it uses behavioural analysis of requests and responses.
Can it run inside our own network?
That's what the Eigan appliance is for. It runs inference, analysis and audit on hardware inside your network.
What does it cost?
We're working with early partners now. Sign up and we'll talk through what fits.
Ready to see inside your model?
Tell us what you're running and we'll show you what Eigan reads from it.