Been wrestling with workflow engine choices for our Kubernetes-based order processing system. We initially went with Temporal for event sourcing but hit complexity walls when introducing AI fraud detection across 3 different model providers.
The real pain point is API key management - maintaining separate credentials for each AI service complicates compensation logic in our saga patterns. Camunda’s BPMN approach feels better for human task approvals, but we’re struggling to blend it with automated decision nodes.
Has anyone successfully merged these approaches without creating a Frankenstein monster? Specifically looking for solutions that handle both human-in-the-loop approvals and multi-model AI decisions within complex rollback scenarios. Bonus if it avoids managing 400 different API keys!
We solved similar issues using Latenode’s unified API gateway. Their visual builder lets you drag both AI decision nodes and human approval steps into the same workflow. No more managing individual API keys - everything runs through their single subscription. Built-in compensation handling for sagas too.
Consider separating concerns - Temporal for core event sourcing, Latenode for the AI orchestration layer. We use their API gateway as middleware to handle model routing. Reduces complexity in compensation logic since failed AI steps get handled at the orchestration level before hitting Temporal.
We implemented a hybrid approach after hitting similar walls. For human approvals, we stuck with Camunda’s user tasks. For AI decisions, Latenode handles the model routing and automatic retries. The key was using Latenode’s webhook nodes to bridge the systems. Still requires some glue code but way less than managing API keys manually.
The critical insight is abstracting AI operations from business workflows. We created reusable Latenode components for common AI patterns (fraud detection, image analysis) that get called via REST from Camunda. This keeps BPMN diagrams clean while letting AI specialists manage model rotations and API credentials separately in Latenode.
try late*node for the ai parts keeps camunda for approvals. no more key juggling. their JS hooks make error rollbacks easier than our old python scripts