Agentic Workflow Automation Platform
Turn plain-language requests into repeatable business workflows
Overview
Instead of asking users to learn menus or scripting, the platform accepts a plain-language description of a task, classifies it against 16 business departments, and lets an orchestration engine compose 18 tools and 30 skills into a concrete workflow. Around 340 workflows are authored across department files, with 60 currently live in the runtime registry, and every generated output goes through a human generate-review-save loop before it becomes a reusable asset.
Problem
- Business teams describe work in natural language, but automation tools demand rigid forms, scripts, or per-task engineering effort.
- Each department has its own vocabulary and output formats, so a single generic chatbot produces answers that are hard to trust or reuse.
- One-off AI generations disappear after use; there was no way to review a result, save it, and run the same workflow again next month.
Solution
- Built a natural-language router that classifies Korean task requests across 16 business departments and hands them to the right workflow family.
- Designed an orchestration engine that composes 18 tools and 30 skills into department-specific workflows, with ~340 workflows authored and 60 live in the runtime registry.
- Wrapped every execution in a human-in-the-loop flow: the engine generates, the user reviews, and approved outputs are saved as reusable business apps alongside CRM, Calendar, and KPI dashboards.
Key Features
Natural-Language Department Router
Classifies a plain Korean request such as 'draft this month's sales settlement email' into one of 16 business departments and selects the matching workflow candidates.
Tool & Skill Orchestration Engine
Composes 18 tools and 30 skills into multi-step workflows, so a single request can chain data lookup, document generation, and delivery steps.
No-code Builder & Marketplace
Lets users create, install, and reuse custom business apps without code, turning proven workflows into shareable building blocks.
Human-in-the-loop Output Flow
Every run follows a generate, review, then save-or-reuse cycle, with results landing in CRM, Calendar, and KPI dashboards instead of a throwaway chat log.
Task-tiered Model Selection
The Node backend routes each task to an appropriately sized LLM via the OpenAI Responses API, reserving stronger models for reasoning-heavy steps.
System Interfaces
Workflow Platform Main Metrics
94.2%
0.8s
Healthy
Process
Mapped recurring tasks across 16 business departments and defined the department taxonomy the router classifies against.
Designed the tool and skill abstractions, then authored ~340 department workflows and promoted 60 of them into the runtime registry.
Built the Node.js orchestration backend with task-tiered LLM model selection on the OpenAI Responses API.
Implemented the React 19 + Vite front end: builder, marketplace, and the generate-review-save output flow with CRM, Calendar, and KPI views.
Hardened the system with Playwright end-to-end tests and packaged the stack with Docker for consistent deployment.
Result
- Shipped an MVP where a plain-language Korean request reliably lands in the right department workflow and returns a reviewable business output.
- Reached ~340 authored workflows across department files with 60 validated and live in the runtime registry.
- Enabled non-developers to assemble and install their own business apps through the no-code Builder and Marketplace.
- Established a repeatable human-in-the-loop pattern where reviewed outputs become reusable assets instead of one-off generations.
Project Specs
- Product Design
- Architecture
- Implementation