Skip to main content

Core Concepts

How it works​

Data owners or organisations deploy the tracebloc client within their own infrastructure, whether on-premises, in the cloud (e.g., VMs), or in Kubernetes clusters running in a DMZ or VPC.

  • The tracebloc client is pulled from DockerHub or GitHub and deployed via Helm or Docker Compose.
  • The data owner prepares and mounts use case specific data and configures compute resources (CPU/GPU, storage) that will be used by external data scientists.

The tracebloc client, once deployed establishes a secure connection to the tracebloc backend over Azure Service Bus (AMQP/WebSocket, port 443). It receives training instructions—such as model configuration, batch size, number of epochs—defined by data scientists.

The tracebloc backend is responsible for:

  • Validating model compatibility, security, and functionality
  • Orchestrating training and managing system stability
  • Aggregating training metrics (e.g., loss, accuracy — but never weights)
  • Providing a Web UI for training feedback, leaderboard access, and diagnostics

Tracebloc enforces strict separation of code and data. Model weights are encrypted and never exchanged between parties, ensuring full differential privacy.

Data scientists upload models and training code via notebooks and interact only with metadata and training metrics. They have no access to raw data at any point.

data privacy
  • Raw data is never shared.
  • Model weights and architecture are encrypted and retained within the data owner's infrastructure. Access is only granted with mutual consent from both parties.