Enables screening LLM input and output through cost-tiered safety checks with per-tenant budgets, returning allow, block, or review decisions. It exposes a JSON-RPC screen tool that isolates tenant spend and falls back to a configurable policy on provider failure or budget exhaustion.
Enables fully air-gapped multimodal document processing by fusing layout parsing with text extraction to produce ordered, labelled reading blocks and reconstructed text via an MCP stdio server.
Enables incremental model scoring over an append-only event log, with a tamper-evident, hash-chained audit trail so every decision can be reviewed and independently re-verified.
Enables graph-based fraud propagation scoring for clinical diagnostics, diffusing risk across entities with a deterministic fallback when the ML model is unavailable.
Enables retail merchandising recommendations via an MCP server that ranks candidates with a Thompson-sampling bandit, tracks per-tenant monthly spend caps, and exposes recommend, record_feedback, and budget_status tools over JSON-RPC stdio.