Private AI Data Flow – Keeping Inference on Your Hardware
Homegrown Intelligence workflows are designed so that model inference, data processing, and result storage occur entirely within the customer's trust boundary. No telemetry, inference payloads, or intermediate tensors are transmitted to external services during normal operation. This report documents the data flow boundaries, encryption mechanisms, and audit capabilities available in the current workflow platform.
Data Flow Boundaries
When a client sends an inference request to a Homegrown Intelligence workflow, the request payload enters the workflow container through a REST or MCP endpoint. The container loads model weights from a local cache directory populated during initial setup. Weights are verified against a SHA-256 manifest before loading; if the manifest check passes, no network access occurs during inference. The processed result is written to a configurable output directory on the host filesystem or returned directly in the HTTP response. The only outbound network activity during normal operation is an optional license check performed once per hour against the local license server on the management VLAN.
Encryption in Transit and at Rest
TLS 1.3 is enforced for all client-to-workflow communication. The reverse proxy terminates TLS using certificates provisioned through Let's Encrypt or a private CA, depending on deployment preferences. Within the internal network, inter-container communication uses mutual TLS with certificates issued by an internal certificate authority. Model weights at rest are optionally encrypted using AES-256-GCM with a key derived from the host's Trusted Platform Module. When enabled, the decryption key never leaves the TPM and weights are decrypted on-the-fly during loading with negligible performance overhead measured at less than 2 percent increase in load time.
Audit Logging and Observability
Every inference request generates a structured log entry containing a correlation identifier, client IP, requested workflow, input size in bytes, processing duration, and output size. Logs are written to stdout in newline-delimited JSON format for ingestion into centralized logging platforms. The audit trail retains entries for ninety days by default with configurable retention periods. A tamper-evident log feature appends a rolling SHA-256 hash chain to each entry, enabling detection of log modification after the fact. Security events such as certificate rotation, configuration changes, and authentication failures are logged to a separate event stream with immediate alert forwarding through the Prometheus Alertmanager integration.
Compliance Considerations
Organizations subject to GDPR, HIPAA, or SOC 2 requirements can configure workflow containers to disable all outbound connectivity except for the license check, which transmits only a hashed machine identifier and workflow version string. The data classification matrix shipped with each workflow package identifies which fields contain personally identifiable information, protected health information, or financial data. A compliance report generator scans the audit log and produces a summary of data access patterns suitable for annual review submissions. The report covers data residence duration, access frequency by principal, and any cross-boundary data transfers that occurred during the reporting period.