Workflow execution
Built 6 node types with conditional routing and external API/LLM execution.
← Back to projects
Case study · AI workflow orchestration
AgentFlow
A multi-tenant AI workflow orchestrator with 6 node types, 4 trigger modes, conditional routing, human approval pause/resume, live GraphQL status, and Gemini execution.
01 · AIGemini LLMAI execution
→
02 · LogicConditionalbranch routing
→
03 · APIHTTP requestexternal integration
04 · HumanApprovalpause / resume
→
05 · DataPersistPostgreSQL state
→
06 · LiveStatusexecution update
GraphQL status keeps execution state visible while the workflow progresses.
The problem
AI workflows need more than a visual builder.
The system needed to execute different kinds of workflow steps, branch conditionally, pause when human approval was required, preserve state, and show execution status while keeping organization access scoped correctly.
What I built
A workflow engine around six node types and four trigger modes.
Human approval
Implemented approval pause/resume so a workflow can stop for a decision and then continue.
Live status
Used GraphQL status updates to make workflow execution observable in the interface.
Multiple triggers
Supported 4 trigger modes for starting workflow execution.
Architecture
Application, identity, GraphQL, workflow state, and AI execution work as one system.
Next.jsworkflow interface
→
Nhostidentity & backend services
→
Hasura / GraphQLdata access & live status
→
PostgreSQL + Geminiworkflow state & AI execution
Engineering decisions
Specific controls are more useful than broad security claims.
Organization-scoped RBAC
Hasura and Nhost are used to keep workflow access scoped by organization and role.
Retries & quotas
Retry handling and quota accounting are part of the execution model rather than being left to the interface.
Webhook security
Webhook security is handled as part of the trigger path for externally initiated workflows.
Proof of work
The project can be reviewed as a live application and as source code.
The live deployment demonstrates workflow creation and execution, while the GitHub repository provides the implementation for the application, backend workflow logic, and supporting configuration.