Inside SafePipe: How We Built a Zero-Knowledge AI Proxy

Inside SafePipe: Architecture of a Zero-Knowledge Proxy

Trust is good, but architecture is better. At SafePipe, we adhere to a strict Zero-Knowledge philosophy. This article explains technically how we process millions of requests without ever becoming a data liability.

1. The "RAM-Only" Guarantee

The most critical aspect of our infrastructure is what we don't do: we don't write request bodies to disk.

When a request hits our Frankfurt Edge nodes:

1. 1The payload is loaded into volatile memory (RAM).
2. 2PII detection algorithms run immediately.
3. 3The request is forwarded to the provider.
4. 4Memory is cleared.

We only log metadata: timestamps, latency, and status codes. The content of your prompt exists on our servers for less than 100 milliseconds.

2. Bring Your Own Key (BYOK) Security

We do not act as a reseller holding a master key. You bring your own API keys (OpenAI, Anthropic, etc.).

• Encryption: Your keys are encrypted using AES-256-GCM before they touch our database.
• Decryption: Keys are decrypted only at the exact moment of a request and only in RAM.
• Isolation: Row-Level Security (RLS) ensures that even if you share a database cluster, your keys are cryptographically isolated.

3. Latency & The Edge

Compliance shouldn't hurt user experience. We utilize Vercel Edge Functions in the eu-central-1 (Frankfurt) region.

• Average added latency: <30ms
• Protocol: HTTP/2 & TLS 1.3 for all connections

4. Multi-Provider Routing

SafePipe supports 8 major AI providers with automatic format conversion:

Supported Models:

• OpenAI: GPT-5.1, o3, GPT-4o
• Anthropic: Claude 4.5 Opus, Claude 3.7 Sonnet
• Google: Gemini 3 Pro
• DeepSeek: DeepSeek-R1, DeepSeek-V3
• xAI: Grok 4
• Meta: Llama 3.3
• Mistral: Mistral Large 2

All requests go through the same PII redaction pipeline before reaching any provider.

Architecture Diagram

┌─────────────────────────────────────────────────────────┐
│              Your Application (EU)                      │
│  ┌──────────┐      ┌──────────────┐                    │
│  │  Web App │ ──── │  SafePipe    │                    │
│  │  Backend │      │  (Frankfurt) │                    │
│  └──────────┘      └──────┬───────┘                    │
└─────────────────────────────┼───────────────────────────┘
                              │
                    ┌─────────┴─────────┐
                    │  PII Redaction    │
                    │  (RAM-only)       │
                    │  < 30ms           │
                    └─────────┬─────────┘
                              │
        ┌─────────────────────┼─────────────────────┐
        │                     │                     │
        ▼                     ▼                     ▼
  ┌──────────┐         ┌──────────┐         ┌──────────┐
  │ OpenAI   │         │Anthropic │         │DeepSeek  │
  │ (US)     │         │ (US)     │         │ (China)  │
  └──────────┘         └──────────┘         └──────────┘

Conclusion

SafePipe isn't just a proxy; it's a firewall. We built it to ensure that even in the event of a catastrophic breach, your customer data remains safe—because we never stored it in the first place.

Practical Implementation

• How to Redact Emails in Node.js
• How to Filter IBANs in Python
• DeepSeek GDPR Compliance Guide

Read the Full Documentation