i build what stays up at 3am

Journeys built in the visual editor compile down to Serverless Workflow specifications that execute on Temporal. Durable execution is the point: a lending flow can span days of applicant back-and-forth and third-party checks, so state lives in the workflow history rather than in a request, and every step carries its own retry policy instead of failing the whole journey.

Each branch can claim a full environment, provisioned onto Kubernetes through Helm charts driven from GitHub Actions. Environments come up with a snapshot of database state, their own domain, and pre/post setup hooks for seeding, which makes them disposable by design rather than long-lived staging that drifts from production.

Third-party behaviour is modelled rather than mocked per test: the simulator reproduces full interaction shapes, including polling responses that settle after N calls and asynchronous webhook callbacks. Integration paths stay exercisable in CI and in ephemeral environments without depending on vendor sandboxes or their downtime.

Events are produced to Kafka to absorb bursts, shaped and routed by Vector, and landed in ClickHouse. The trade is deliberate: giving up row-oriented storage buys the compression and scan performance that high-cardinality event data needs, which is what makes wide aggregate queries over log volume affordable.

Products hand off events rather than implementing their own delivery. Firefly queues them on AWS SQS to decouple production from delivery, and drives a configurable retry and backoff schedule through AWS EventBridge so a slow or flapping consumer stalls only its own deliveries. At-least-once is the explicit contract, which pushes idempotency to the receiver.

Configuration is treated as versioned data with an audit trail, so any change can be rolled back to a known revision instead of being hand-reverted. Updates propagate to distributed consumers without a restart, and access is exposed over gRPC behind fine-grained role-based access control so tenants and teams only reach their own surface.

Rather than pinning a vendor in config and retrying into a degraded one, each request is routed using live latency, success-rate, and cost signals. Degradation becomes a routing decision instead of a retry storm. Signals are aggregated in Redis for fast reads with PostgreSQL holding the vendor and policy definitions.