What runs where, exactly?
Three field tiers, one discipline. A Gateway Node does integration and privacy only: EHR bridging, PII scrubbing at the source, a disk-persisted offline queue, a hardware-bound VPN uplink, and no AI inference. A Clinical Node adds local triage and pre-screening models, imaging ingest, and vocal-biomarker telemetry. An Enterprise Node runs the complete consilium, the graph knowledge layer, and the digital-twin engine entirely on premises, with a fully autonomous air-gap mode. Hardware is commodity, from single-board gateways to workstation-class nodes.
How do you know the model on our site hasn’t been tampered with?
It refuses to start. At boot, every model artifact is hashed (SHA-256) and matched against a signed central registry; packages are additionally signed with DeepSensi’s private key (RSA-2048) and verified against a public key held in the device’s secure enclave. On any mismatch the node fails closed: initialization blocks, local execution locks, and a high-priority security alert fires. Aligned with FDA SaMD guidance and IEEE P2941 model-integrity principles.
How does DeepSensi connect to Epic or Cerner?
Through standards, not custom glue: SMART on FHIR sign-in (OAuth 2.0 with PKCE), reading Patient, Observation, and DiagnosticReport resources; the finalized result is written back as a FHIR ClinicalImpression plus a digitally signed DiagnosticReport, under the physician’s signature. Epic, Cerner, Allscripts, and OpenEMR are supported through the same connector suite.
What actually leaves the building?
Nothing identifiable. A local, deterministic scrubber redacts names, dates, geographic identifiers, and serial numbers under HIPAA rules before any context leaves the local network; research cohorts additionally receive k-anonymity (k ≥ 5) and l-diversity treatment. On an air-gapped enterprise node, the answer is simpler still: nothing leaves at all.
What happens when the network fails?
The node notices before you do. An uplink monitor watches link quality continuously; past tolerance, it switches to a degraded state: reasoning falls back to local on-device models, a context-compression engine strips non-essential history to save bandwidth, and timeout budgets extend gracefully. Satellite uplinks are first-class citizens. Clinics keep working when the link does not.
How is clinical data separated from billing?
By hard account isolation on the internal event mesh: clinical channels and financial channels are logically segregated planes, so a compromise of one cannot traverse into the other. Failed messages land in dead-letter queues for diagnostics; throughput scales elastically to hundreds of MB/s for imaging and population-scale simulation.
How does federated learning work here?
Gradients travel; records never do. Nodes train locally on hospital data and export only weight adjustments to the central repository. Combined with source-level de-identification and cohort anonymization, this is how pharma-grade research runs without a single patient record changing hands.
Why is the audit trail legally defensible?
Every step of the reasoning chain, prompts, agent interactions, consensus changes, is cryptographically signed (RSA-2048) into an immutable ledger, and the session log’s root hash is periodically anchored to a public Layer-2 blockchain, an anchoring mechanism that is patent-pending. The trail is tamper-evident even against the operator: court-grade provenance for liability scenarios.
Which security standards do you implement?
TLS 1.3 with mutual authentication and certificate pinning, hardware-bound VPN tunnels, secure-enclave key storage, boot-time model attestation, two-factor authentication per RFC 6238, HIPAA and GDPR compliance, and EU AI Act alignment by architecture. The deeper engineering dossier, model inventory, mesh topology, performance envelopes, is available under NDA.
Public edition. Model inventory, mesh topology, performance envelopes, and deployment runbooks: under NDA via [email protected].