Back to projects

Wikit Semantics Workflow

Durable Workflow Engine for Conversational AI

Distributed SystemsBackend EngineeringTypeScriptWorkflow OrchestrationEvent-Driven ArchitectureConversational AILLM ApplicationsTemporal
Depiction of an AI workflow

Wikit needed a way to run complex conversational AI flows reliably—beyond fragile prompt chains or single-service backends.

I designed and built a workflow engine that turns AI interactions into structured, durable executions with state, branching, and real-time feedback.

A single conversation can branch on user intent, invoke external services, retry failed steps, and stream intermediate progress back to users—making AI behavior more resilient, observable, and production-ready.

Core Capabilities

  • Workflow engine for multi-step conversational AI with stateful execution
  • Conditional branching and parallel step orchestration
  • Real-time event streaming for live progress feedback
  • Built-in retries, failure handling, and execution control
  • Extensible plugin system for adding new workflow steps

Key Results

  • Replaced fragile prompt chains with durable workflows (state, retries, controlled execution lifecycle)
  • Improved observability through real-time events and persistent activity logs
  • Enabled safe iteration on AI behavior with clear execution traces and debugging visibility
  • Reduced development friction via standardized configuration and tooling across the monorepo

Architecture Overview

Wikit Semantics Workflow is a TypeScript-based workflow platform built on Temporal, structured as loosely coupled services:

  • Gateway (API + Event Streaming)
    Handles external requests and streams workflow events to clients in real time
  • Workflow Engine (Orchestration)
    Manages workflow state, branching logic, and execution control using Temporal
  • Worker (Activity Execution)
    Executes individual workflow steps (AI calls, external services, transformations)

Technical Highlights

  • Modelled workflows as directed graphs with typed schema validation
  • Implemented branch pruning and parallel step scheduling
  • Built a plugin-based activity registry for extensibility
  • Designed a Redis-backed event layer separating real-time signals from persistent logs
  • Standardized TypeScript configuration, linting, and tooling across a monorepo

My Role

Owned the system end-to-end:

  • Designed architecture and workflow model
  • Implemented the runtime and orchestration layer
  • Built service infrastructure and event system
  • Developed shared tooling for consistency across apps and packages
Website