Metis Security Overview¶
For client security reviews and compliance conversations.
Data Handling Principles¶
- Your data stays yours. We never use client data to train models, share it with other clients, or retain it beyond the engagement.
- Tenant isolation. Each client's documents are stored in a completely separate, encrypted namespace. No cross-client data access is possible.
- Minimal retention. Data is deleted or returned within 30 days of engagement termination.
- Transparency. We disclose exactly how your data is processed and which services touch it.
Infrastructure Security¶
| Control | Implementation |
|---|---|
| Encryption in transit | TLS 1.3 via AWS ALB + ACM certificate |
| Encryption at rest | AES-256 server-side encryption on all S3 storage |
| Access control | IAM role-scoped access, no public bucket access, admin token authentication |
| Tenant isolation | Each client has a dedicated, isolated storage namespace in S3 |
| Network security | AWS VPC with security groups, ALB-only ingress |
| Container security | Minimal Docker image (python:3.11-slim), no root access, health-checked |
| Secrets management | API keys and tokens stored in AWS SSM Parameter Store (SecureString) |
| Logging | CloudWatch with 30-day retention, no PII in application logs |
| Infrastructure as code | All infrastructure defined in Terraform, auditable and reproducible |
Data Processing Flow¶
Your Documents → Encrypted Upload → Chunking & Indexing → Encrypted Storage (S3)
↓
User Question → Retrieval (within your namespace) → LLM Synthesis → Answer
What happens to your documents¶
- Upload: Documents are transmitted over HTTPS and stored in your isolated S3 namespace with AES-256 encryption
- Processing: Documents are split into searchable chunks, embedded as vectors, and indexed. All processing happens within AWS infrastructure
- Storage: Vector indexes and metadata are stored encrypted in S3. Original document text is preserved only in chunk form within the index
- Retrieval: When a question is asked, only the most relevant chunks (typically 5 passages) are retrieved from your namespace
- Synthesis: Retrieved passages are sent to the LLM provider for answer generation
- Deletion: Upon termination, all client data (indexes, metadata, uploads) is permanently deleted from S3
LLM Processing¶
Answer synthesis uses an external LLM API. The current default is Groq (Llama 3.1 8B). Only the retrieved document chunks relevant to each query are sent to the LLM — not your entire document library.
For clients requiring data to remain within controlled infrastructure, we offer: - AWS Bedrock — LLM processing stays within our AWS account. No data leaves AWS infrastructure. - Self-hosted models — LLM runs on dedicated compute within our infrastructure. Zero external data transmission.
These options are available for full deployment and enterprise engagements.
What We Don't Do¶
- We never use your data to train AI models
- We never share your data with other clients
- We never retain your data after the engagement ends
- We never allow cross-tenant data access
- We never store passwords, payment information, or authentication credentials from your documents
Compliance Readiness¶
| Framework | Status |
|---|---|
| SOC 2 Type II | AWS infrastructure is SOC 2 certified. Application-level audit available on request. |
| HIPAA | Available for enterprise deployments with AWS Bedrock + BAA. Contact us for details. |
| GDPR | Data processing agreement available. Data residency in US-East (Virginia). |
| PCI DSS | Not applicable — Metis does not process payment card data. |
Recommendations for Clients¶
For pilots and evaluations¶
- Provide operational documents (procedures, policies, manuals) rather than documents containing individual PII
- We sign mutual NDAs and data processing terms before receiving any documents
For full deployments with sensitive data¶
- Request AWS Bedrock or self-hosted LLM configuration (no data leaves AWS)
- We can implement PII detection and redaction in the retrieval pipeline
- Customer-managed encryption keys (SSE-KMS) available on request
Questions?¶
Contact us to discuss your specific security and compliance requirements.
[CONTACT_EMAIL] [CONTACT_PHONE]