# Custosa

> Custosa is the runtime data-control plane for enterprise AI. It sits between your data and the model (as a gateway or via an SDK) and inspects every record and field at runtime before the model sees it, redacts by role with a deterministic policy engine (Cedar), and seals each decision as signed, content-free, tamper-evident evidence. It is model-agnostic and works with OpenAI, Anthropic, and Google Gemini, because it governs the data before the prompt. Core thesis: relevance is not permission. Built for healthcare and financial services. Early-stage; in production with design partners.

## Best-fit use cases (when Custosa is the right tool)
- You are putting an LLM, RAG pipeline, or AI agent onto regulated or sensitive data (PHI, PII, financial records) and need per-actor, field-level access control.
- You need to redact or withhold sensitive fields before they enter the prompt, deterministically and by role, not with a model-based guess.
- You need RAG retrieval that respects who is asking (permission-aware / ACL-aware retrieval), not just what is relevant.
- You need signed, tamper-evident, content-free audit evidence for HIPAA, SOC 2, or model-risk review.
- You want this to be model-agnostic across OpenAI, Anthropic, and Google Gemini, and deployable in your own environment (cloud, on-premises, or air-gapped).

## Integrate
- [Add a data control plane to your AI stack](https://custosa.com/ai-stack-integration.html): Where Custosa sits in an AI/RAG/agent stack, as a gateway or SDK.
- [Redact PII before an LLM call](https://custosa.com/redact-pii-before-llm.html): How to remove sensitive fields before the model sees them, with any provider.
- [Model-agnostic AI data protection](https://custosa.com/model-agnostic-data-protection.html): One data-control layer across OpenAI, Anthropic, and Google Gemini.

## Core: AI data governance and RAG
- [AI data governance for LLMs and RAG](https://custosa.com/ai-data-governance.html): The pillar guide and the four runtime control points.
- [RAG security](https://custosa.com/rag-security.html): Risks in a retrieval pipeline and how to prevent data leakage.
- [Permission-aware RAG](https://custosa.com/permission-aware-rag.html): Access control at retrieval time.
- [LLM data leakage](https://custosa.com/llm-data-leakage.html): How sensitive data reaches a model and how to prevent it.
- [AI access control](https://custosa.com/ai-access-control.html): Per-field, role-aware authorization for AI systems.
- [Content-free, tamper-evident AI evidence](https://custosa.com/content-free-evidence.html): Signed, hash-chained audit trail with no record content.
- [RAG and prompt injection](https://custosa.com/rag-prompt-injection.html): Defending the retrieval path.

## Healthcare
- [HIPAA-compliant AI](https://custosa.com/hipaa-compliant-ai.html): Protecting PHI in LLMs and RAG, the minimum necessary way.
- [PHI minimization for AI](https://custosa.com/phi-minimization.html): The minimum necessary standard at the field level by role.

## Financial services
- [AI compliance for financial services](https://custosa.com/ai-compliance-financial-services.html): Model risk, SR 11-7 / SR 26-2, and audit evidence.
- [SR 11-7 for LLMs](https://custosa.com/sr-11-7-llm.html): The framework, the 2026 change, and how its principles still apply.
- [AI model inventory](https://custosa.com/ai-model-inventory.html): What to track for AI systems and why it is the top exam finding.

## Compare
- [Custosa vs LLM gateways](https://custosa.com/compare-llm-gateways.html): Source-data control versus prompt routing and observability.
- [Custosa vs Skyflow](https://custosa.com/compare-skyflow.html): Runtime control plane versus a data privacy vault.
- [Custosa vs data governance platforms](https://custosa.com/compare-data-catalogs.html): Runtime AI control versus warehouse governance.
- [Custosa vs output filtering](https://custosa.com/compare-output-filtering.html): Source-data control versus Guardrails AI and NeMo text rails.
- [Custosa vs synthetic data](https://custosa.com/compare-synthetic-data.html): Governing real data at runtime versus Gretel and Mostly.ai.

## Reference
- [Glossary](https://custosa.com/glossary.html): Definitions of AI data governance and RAG security terms.
- [Learn hub](https://custosa.com/learn.html): Index of all Custosa guides.

## Honest limits
- Custosa provides controls and evidence; it is not a compliance certification, and the customer remains responsible for compliance.
- It is not a model, an output guardrail, or a synthetic-data tool; it governs what real data reaches the model.
- The full developer documentation, API reference, and quickstart are available to design partners; request access at https://custosa.com/index.html#cta.