Security and data handling
Five questions to ask about any AI on your program.
We design every system against these five questions, the AI Assurance layers. None of them is ever finished. Ask them of us, and of anyone else bidding the work.
01
Sovereignty
Are we building our own AI, or renting someone else's?
We can build without receiving your sensitive data. With a defense prime, we built an intelligent knowledge graph over engineering documentation, without receiving CUI. We built it; the prime tested it. Our codebase assessment reads code and architecture, never the data your systems hold.
02
Security
What can our AI reach, and what could it open up?
LLM-light: mostly deterministic, with constrained and checked LLM steps. We start from a secure foundation and box the generative model so it can't open a hole. Our systems are designed to run where your data lives, on open-weight or frontier models, in enclaves, or air-gapped.
03
Governance
Who checked what our AI produced, and what are we putting into AI tools?
Every output traces to its source document and page. When sources conflict, we keep both and show where each came from, so your experts can resolve it. Low-confidence output goes to a person before it moves on.
04
Intent
Does our AI give the answer we meant, the same way every time?
We move decisions from open-ended generation to rules and classifiers, and check that identical runs give identical results. No model is right every time, so we design for answers that are consistent, visible and correctable.
05
Cost
What are we spending on AI, and what is it returning?
We cap tokens, route each task to the right model, and build small purpose-built models where they cut compute. We size the system to the hardware your program actually has.
Posture
CMMC Level 2 (self-assessed)
NIST SP 800-171 implemented
FISMA Moderate control alignment
SOC 2 operational experience
Designed for GovCloud, Azure Gov and air-gapped operation