Architecture & Security
Understanding how your data is processed is critical for security and privacy. ModelBeam’s architecture is designed for high-performance AI inference with robust data protection at every layer.1. Security Architecture
The following outlines the data flow for a standard inference job.Data Flow
- Client → API (HTTPS Encrypted) — Requests are initiated via the REST API. All data transmission is encrypted in transit.
- API → Worker — The API validates the request, deducts credits, and dispatches the job to an isolated GPU worker with the model parameters and input data.
- Worker (Processing) — The GPU worker receives the payload (images, text, audio). Data is processed in an isolated execution environment with no persistent storage.
- Result → Client — The generated asset is uploaded to encrypted storage and a download URL is returned to the client via the same encrypted channel.
Security Mechanisms
Multi-layer security measures are implemented to guarantee integrity and data protection:- Encryption Everywhere — Encryption is enforced at every stage of communication. All API endpoints, WebSocket connections, and storage URLs are encrypted in transit.
- Authenticated Callbacks — Worker-to-API callbacks are authenticated to prevent spoofing or injection of fraudulent results.
- Input Validation — All inputs are validated for type, size, and content before processing. File uploads are checked against size limits and allowed formats.
- Rate Limiting — Per-endpoint rate limits protect against abuse and brute-force attacks.
- Injection Protection — External URLs and references are validated against allowlists. Internal network access is blocked.
2. Privacy Model
Data Visibility
| Component | What it sees | Retention |
|---|---|---|
| API Server | Request metadata, prompts, parameters | Job records stored for billing and history. No long-term storage of input payloads. |
| GPU Worker | Full payload (images, audio, text) | Data exists only in memory during processing. Cleared after job completion. |
| Result Storage | Generated outputs only | Stored with server-side encryption at rest. Available via download URLs. |
Additional Safeguards
- Secure Token Storage — Authentication tokens are protected against client-side theft.
- Encryption at Rest — All generated results are encrypted at rest in storage.
- No Logging of Sensitive Data — Prompts, images, and generated content are never written to application logs.
- Webhook Signatures — Outbound webhooks are signed, allowing recipients to verify authenticity and prevent replay attacks.
3. Use Cases
ModelBeam is a strong fit for:- Public content generation — Images, videos, speech, and music
- Content transcription — YouTube, X, TikTok, Twitch, Kick
- AI workloads without critical data — Creative tools, prototyping, demos