qbrixHow qbrix Works
qbrix separates the hot path (selection) from the learning path (training) to achieve lightning-fast decisions with continuous parameter updates.
Select
Client sends a request to proxysvc. Feature gates are evaluated, then the request is routed to motorsvc via gRPC for instant arm selection.
Feedback
Reward signals are published to Redis Streams via proxysvc. Durable event sourcing ensures no feedback is ever lost, even during traffic spikes.
Train
cortexsvc consumes feedback in batches, trains bandit algorithms, and writes updated parameters to Redis — ready for the next selection.
Adapt
motorsvc reads fresh parameters from Redis with TTL-based caching. Decisions improve continuously without any downtime or redeployment.
The Core Idea
By separating the hot path (selection) from the learning path (training), qbrix delivers ultra-low latency decisions while models continuously improve in the background. No tradeoff between speed and intelligence.