Live demo scenarios5 reproducible security proofs: No provider API keys in agent pods • gVisor Python sandbox • Cryptographic tenant isolation • Gateway request circuit breakerExplore the 5 Proofs
Solutions · Healthcare & Life Sciences

Keep clinical and research agents inside strict tenant, network, and log-redaction boundaries.

Run LangGraph, LangChain, and Deep Agents workloads with prompts and keys kept out of default logs, self-hosted Langfuse traces that never leave your cluster, verified JWT tenant isolation, and gVisor-isolated Python execution.

Risks

Where it fails without a runtime boundary

Patient Context in Container Logs

Standard stdout logging accidentally capturing clinical prompts or auth tokens. How Zelkor stops it: Prompts and keys stay out of default logs and trace metadata.

Traces Sent to External Cloud Vendors

Shipping execution traces containing research workflow data to a third-party SaaS. How Zelkor stops it: Langfuse runs self-hosted inside your cluster on ClickHouse, Postgres, and S3-compatible storage.

Cross-Study Data Bleed

Multi-site clinical trial assistants querying records from the wrong institution or cohort. How Zelkor stops it: Verified JWT identity binds every Postgres query, Qdrant search, and S3 path to the caller’s tenant.

Unisolated Biostatistical Python

Agent-generated statistical scripts running inside the main application container. How Zelkor stops it: Untrusted Python runs over MCP in a separate warm pool of gVisor workers.

Unchecked Clinical Prompt Trajectories

Off-topic or unsafe completions reaching clinical staff. How Zelkor stops it: NVIDIA NeMo Guardrails inspects chat completions and fails closed when enabled.

Leaking Internal Tokens to External Tools

Forwarding hospital JWTs when querying external literature MCP servers. How Zelkor stops it: External MCP backends authenticate with dedicated Kubernetes Secrets; the caller’s JWT is stripped.

Controls

Compliance & Control Mapping

RequirementWhat Zelkor DoesHow You Check It
Log & Telemetry MinimizationOmits raw prompt bodies, bearer tokens, and API keys from default container logs and trace metadata.Inspect pod stdout logs and Langfuse span attributes.
In-Cluster Audit ResidencyStores all traces and automated run scores in self-hosted Langfuse inside your Kubernetes cluster.Verify cluster storage claims and outbound NetworkPolicy.
Cohort & Institution SeparationVerifies front-door JWTs and rejects any model-supplied tenant ID on MCP tool calls.Run Two Desks, One Graph cross-tenant proof.
Isolated Scientific ComputeRuns generated Python inside gVisor worker pods that require an internal gateway token.Run The Shell Moved sandbox proof.
Topologies

3 Healthcare & Life Sciences Workflows

Multi-Site Clinical Trial Assistant

Query trial metadata through read-only Postgres with site-level Row-Level Security and search protocol documents in Qdrant without cross-site leakage.

Sandboxed Biostatistics Execution

Run survival analysis and data normalization scripts in the warm gVisor worker pool while keeping the agent pod isolated.

Governed Literature Search

Connect agents to external research MCP servers using gateway-managed Kubernetes Secrets without exposing internal clinician tokens.

Run clinical and research agents with self-hosted traces and hard isolation boundaries.