qbrixqbrix

Introduction

qbrix is a cloud platform for multi-armed bandit (MAB) optimizations. It separates the hot path (arm selection) from the learning path (training) to achieve ultra-low latency decisions with eventual consistency in parameter updates.

Architecture

qbrix is composed of three services, two data stores, and a message queue:

qbrix architecture

ServiceRole
proxysvcGateway — HTTP/REST + gRPC entry point, auth, experiment management, feature gates
motorsvcSelection — stateless, horizontally scalable, reads cached params
cortexsvcTraining — single instance, batch training via Redis Streams

For a deep dive into service internals, caching layers, and performance optimizations, see Architecture.

Key Concepts

ConceptDescription
PoolA named collection of arms (variants). For example, a pool of homepage hero images.
ArmA single variant within a pool. Each arm has an index, a name, and optional metadata.
ExperimentLinks a pool to a policy (algorithm). Defines how arms are selected and trained.
PolicyThe bandit algorithm used for selection. qbrix ships 12 policies across stochastic, contextual, and adversarial categories.
Feature GateControls experiment traffic with rollout percentages, scheduling, and targeting rules.
FeedbackA reward signal sent after selection. Used by cortexsvc to update policy parameters.
ContextOptional request-level information (user features, metadata) used by contextual policies.

Request Flow

  1. Sign up at cloud.qbrix.io and create an API key
  2. Create a pool with arms via the API or console
  3. Create an experiment linking the pool to a policy
  4. Integrate select and feedback API calls in your application
  5. Monitor and optimize via the console dashboard

What's Next