Findola Platform · Sovereign / edge
Findola for the networks that cannot touch the cloud. Sovereign AI search on-prem and air-gapped.
Vaulta is Findola's sovereign deployment. The full search and agent stack ships onto on-prem NVIDIA hardware so hospitals, banks, defense, and public-sector teams get cited answers and agentic action without data ever leaving their network. It is the appliance that opens buyers cloud-only competitors cannot serve.
Why Vaulta exists.
The highest-value, most document-heavy organizations in healthcare, finance, defense, and government are legally or operationally barred from sending data to cloud AI. They have the worst search problem and no compliant option. SaaS-only tools are non-starters in SCIFs, hospitals, and regulated data centers.
What it does
Vaulta runs Cortexa, Retrova, and Agentiq as self-contained NIM microservices on Jetson and IGX-class hardware, or the customer's own on-prem GPUs, through NVIDIA AI Enterprise. It ships as a hardened, single-tenant appliance with local indexing, local inference, BYOK encryption, and SIEM audit export, with zero outbound dependencies. The same Findola experience, fully sovereign.
What ships in Vaulta.
Every capability below maps to a real workload that needs accelerated compute.
Single-tenant appliance
Hardened, with zero outbound calls.
Local everything
Cortexa graph, Retrova retrieval, and Agentiq agents run on-device (NIM).
Enterprise controls
BYOK encryption, SSO/SCIM, RBAC, and SIEM audit export.
Offline updates
Model updates arrive as signed, air-gap-friendly packages.
Compliance packs
HIPAA, FedRAMP, ISO 27001/42001, GxP, and the EU AI Act.
Architecture.
Customer data stays local. Connectors feed a local CDC pipeline into an on-appliance GPU index (Cortexa), then Retrova retrieval and Agentiq actions, all as NIM services on Jetson or IGX or the customer's GPUs under NVIDIA AI Enterprise. Holoscan handles local stream throughput. Updates are signed bundles, and nothing phones home.
NVIDIA hardware
- IGX Orin / Jetson Orin. Air-gapped edge inference in a sovereign appliance.
- On-prem H100 / L40S. Scale inside larger private data centers.
- Holoscan dev kits. Sensor and stream ingestion for live local indexing.
NVIDIA SDKs & libraries
- Holoscan. High-throughput local ingestion to keep the index live.
- TensorRT. Compile models to run at low latency on constrained edge GPUs.
NVIDIA software
- NVIDIA AI Enterprise. A supported on-prem platform regulated buyers require.
- NIM. Packaged microservices, the same stack, air-gapped.
Vaulta is GPU-essential.
Local inference for retrieval, reranking, and agents has to run on the appliance's GPUs, with no cloud offload allowed. TensorRT plus Jetson and IGX make sub-2s on-device answers feasible. CPU-only edge cannot run the model stack at usable latency.
The moat.
The on-prem footprint, compliance accreditation, and local data gravity create very high switching costs. Once Vaulta is inside the secure network, it stays. It reuses the owned models from Lumind.
Vaulta adds a physical/edge-AI and sovereign-AI dimension on NVIDIA hardware (Jetson, IGX, Holoscan, AI Enterprise), exactly the strategic areas NVIDIA prioritizes, and a moat reviewers respect.
Ecosystem link: Vaulta is the sovereign deployment surface. It is how Cortexa, Retrova, and Agentiq reach regulated buyers, using owned models from Lumind.
See Vaulta on your own data.
Connect three apps and Findola will answer your first question in under ten minutes.