Real-time IoT analytics with autonomous agent architecture, Kafka-backed stream processing, and microsecond circuit breaker protection against cascade failures.
Real-time systems fail in one of two ways: they stop working, or they keep working while producing wrong results. The second one is worse. Silent failures at high throughput mean thousands of bad records downstream before anyone notices.
Phronis was built around one constraint: a downstream failure cannot be allowed to cascade. If Grafana is slow, the pipeline cannot back up. If an agent throws, recovery must be automatic, logged, and traceable to a specific message offset.
The circuit breaker pattern, borrowed from electrical engineering, is the mechanism. When a consumer exceeds the failure threshold, the breaker trips open. Traffic routes to the dead letter queue. The remaining consumers keep running. The operator gets an alert in under 600ms.
From sensor event to operator alert. Every millisecond is measurable. Every stage logs to a named Kafka topic. The breaker state is queryable via FastAPI at any point in the chain.
Every capability in Phronis is a @tool. Every agent is a @agent. The decorator pattern keeps the circuit breaker logic orthogonal to the business logic. Agents can be composed, swapped, or extended without touching the CB implementation.