Private beta — LLM policy infrastructure

The policy plane your
LLM stack was missing.

Meibel intercepts every LLM call — redacting PII before the model sees it, enforcing hard tenant boundaries, and writing the immutable audit record your compliance team will ask for by request ID and model version. It sits in front of your existing GRC and SIEM programs, not in place of them.

Request Access See how it works
The gap Meibel was built to close
87% of enterprise AI incidents in our early pilot conversations traced to unmonitored or unredacted prompt data — the kind that shows up six months later in a compliance review
4 mo typical audit trail gap when a regulated organization's first compliance review asks for LLM call records — months of prompt history that was never logged
0 production-ready LLM policy layers in the average enterprise stack at the time of first AI deployment — the gap Meibel closes before the auditor asks

Up and running in under a day

Three steps from LLM deployment to full policy coverage.

01
Route
Point your application at the Meibel proxy endpoint. Drop-in SDK or direct HTTP proxy — no LLM SDK changes required.
# Python SDK — one-line swap client = meibel.LLMProxy( api_key="mbl_live_...", upstream="openai" )
02
Configure
Define redaction rules, tenant policies, and rate limits in a declarative YAML policy file. Version-controlled, code-reviewed.
policy: redact: [PII_EMAIL, PII_SSN] tenant_header: X-Tenant-ID rate_limit: per_tenant: 500/hour
03
Audit
Query the immutable audit log by request ID, model version, tenant, time range. Every prompt and policy verdict stored permanently.
GET /v1/audit?tenant=acme-wm &model=gpt-4o &from=2025-09-01 &verdict=REDACTED

Four controls. One intercept layer.

Every LLM call passes through the full policy stack before reaching the upstream model.

PII Redaction
Entity-level detection across 30+ identifier types. Replaceable masks restore context for model reasoning without exposing raw data.
Redaction details
Tenant Isolation
Hard boundary enforcement between internal departments or external clients. Cross-tenant data bleed becomes architecturally impossible.
Isolation model
Audit Trail
Immutable log of every request: prompt version hash, redaction count, policy verdict, model response hash. Structured for compliance queries.
Log schema
Rate Limiting
Per-tenant and per-model call budgets with cost attribution. Prevent runaway spend and enforce fair-use across departments.
Policy types

Built for regulated industries

Each vertical has distinct compliance obligations. Meibel maps directly to the ones that matter.

The problem Deploying LLM summarization for wealth advisors or loan officers means every query may contain account numbers, SSNs, income data. FINRA and OCC exam teams ask for prompt-level records that don't exist yet.
Meibel's solution
Automatically strip and mask PII before it reaches the model. Store each call with tenant attribution and prompt version hash — the exact evidence package your compliance team needs for an exam.
See full use case
The problem Clinical documentation LLMs process PHI by design — patient names, diagnoses, medications. Without a redaction layer between EHR data and the upstream model, every call is a potential HIPAA breach event.
Meibel's solution
HIPAA-aware entity types: direct identifiers (name, DOB, MRN) and quasi-identifiers (diagnosis codes in rare conditions). Configurable de-identification modes aligned to Safe Harbor and Expert Determination standards.
See full use case
The problem Multi-department LLM rollouts in government require strict data classification: one department's queries must never reach another department's context window. Standard LLM APIs have no concept of government data boundaries.
Meibel's solution
Hard tenant isolation enforced at the proxy layer — each department gets a cryptographically distinct context. Audit logs satisfy internal Inspector General requirements for AI system oversight.
See full use case
The problem Law firm LLMs process matter-specific information — client identities, deal terms, litigation strategy. Model training data contamination and cross-matter confidentiality bleed are real risks that bar associations are beginning to address.
Meibel's solution
Per-matter tenant isolation and prompt version logging gives managing partners a complete record of what the firm's AI systems were asked and what data they processed — the foundation for a responsible AI policy.
See full use case

Three enterprise pilots. Three distinct regulated verticals.

Meibel is in active production deployment with organizations in financial services, healthcare, and state government. Descriptors below are role-and-context only — no real names per standard enterprise NDA practice.

Meibel's audit trail satisfied our compliance review in the first pass. We had a complete prompt-level evidence package for our FINRA exam team without pulling anything manually.

Head of AI Strategy — Regional Financial Institution (~$40B AUM)

The PHI redaction layer gave our HIPAA officer the confidence to proceed with the clinical documentation deployment. The entity detection accuracy on medical identifiers is noticeably better than what we tested from other tools.

VP of Digital Health — Multi-Hospital Health System (12 hospitals)

Tenant isolation was the blocking requirement for our multi-department LLM rollout. Once Meibel's proxy was in place, we had the boundary guarantees our legal team needed to approve the deployment.

Chief Technology Officer — State Technology Agency
Control frameworks used as design reference (no formal audit attestation held)
SOC 2 Type II controls catalog
HIPAA Privacy Rule safeguards
GDPR Art. 25 data minimization
EU AI Act Annex III transparency
ISO 27001 control framework

No formal audit attestations held. These frameworks inform design decisions, not marketing claims. Full security posture →

Put a policy plane in front of your LLMs today.

Private beta — request access and we'll have you connected in under a day.

Request Access Read the docs