A Levenhall Product
AI evaluation & governance for the enterprise.
The substrate for AI insurability.
AI is now deployed inside every regulated industry. Underwriters, risk committees, and compliance officers have no standardized way to determine whether a deployment is insurable, defensible, or compliant.
The deployments don't stop while they figure it out.
AIGIS exists to answer one question, continuously: is this model safe and mission-ready for this user, on this network, under this regulatory regime?
Not a research bench. Not a one-time audit. A continuous assessment substrate that AI insurance, regulatory compliance, and enterprise risk all run on top of.
Insurers cannot price what they cannot measure. AIGIS produces the continuous evaluation record that turns an AI deployment into a measurable risk: every run captured with a full environment snapshot, every adversarial probe logged against MITRE ATLAS, every degradation event reproducible, every model and prompt versioned.
For underwriters: a continuous signal to price against. For brokers: a defensible record at renewal. For the enterprise: the compliance evidence the insurer needs without a fire drill every quarter.
Vendor-agnostic. Identical container images for managed cloud, segregated networks, and air-gapped environments. The only thing that changes is a Helm config.
llm: provider: openai model: gpt-4o vector: backend: managed-cloud network: egress: full
llm: provider: azure-openai model: gpt-4o vector: backend: pgvector network: egress: restricted
llm: provider: ollama model: llama3-70b-q4 vector: backend: pgvector network: egress: none
Open-source underneath. No vendor lock-in at the layer that matters. The same image runs in your AWS account, in a segregated VPC, or fully air-gapped against a local Ollama model.
One uniform protocol behind adapters for OpenAI, HuggingFace, Bedrock, gRPC, and local Ollama. Pluggable scorers covering exact match, similarity, classification, and routed human evaluation — with Krippendorff's alpha and Fleiss' kappa for inter-rater reliability. Every run captures a full environment snapshot so a result is reproducible months later.
An automated red-teaming pipeline aligned to MITRE ATLAS — prompt injection, jailbreak generation, data-poisoning probes, model-extraction resistance — with results mapped to ATLAS technique IDs an insurer or auditor recognizes. An operational-stress simulator measures degradation curves under latency, jitter, and throttling, not binary pass/fail.
Live coverage against the frameworks insurers and regulators actually use. Risk scoring against active incidents. Insurance coverage monitoring across the portfolio. The output is a defensible governance trail, refreshed continuously, not a snapshot taken once a year.
Levenhall Proprietary
We work with AI insurance counterparties, regulated enterprises, and institutional risk teams on early access deployments.