Core Technology

AI Agent Orchestration Platform

Not just an LLM wrapper — a production-grade agent lifecycle management platform that enforces accuracy, consistency, and full observability across every AI agent, regardless of domain.

Accuracy & Consistency

AI output always matches predefined JSON Schema — hallucination structurally prevented

Version Control

All prompts managed via CI/CD — rollback available at any time

Explainable AI (XAI)

Every pipeline stage logged and visualized in real time via Live Broadcast Mode

14+

Agents in Production

Domains Deployed

100%

Schema-Validated

Pipeline Stages

Overview

Platform Overview

The AI Agent Orchestration Platform is the common backend engine powering all AI features across MSoftech's products. Rather than building AI logic per-project, every agent — regardless of domain — runs through the same standardized pipeline.

The platform's core challenge is the gap between what LLMs produce (free-form text) and what applications need (structured JSON). The solution is a 3-artifact architecture — Registry, Template, Schema — where each artifact is independently versioned and CI/CD managed, eliminating the need for code deployments when prompts change.

With Live Broadcast Mode, every pipeline stage is streamed in real time — showing exactly which prompt template was used, how data was injected, what the LLM received, and whether the output passed schema validation. This makes AI behavior fully traceable and auditable.

AI Engine

Multi-Provider — Gemini / GPT-4

Deployed In

4 Projects · 8 Agents

Core Guarantee

100% Schema-Validated Output

Pipeline

5-Stage Core Pipeline

모든 에이전트가 동일하게 통과하는 표준 흐름

① Request Intake

Client sends data + request info

② Prompt Generation

Registry lookup + data injection

③ LLM Call

Send assembled prompt to AI provider

④ Schema Validation

100% schema match check

⑤ Result Delivery

Validated result to client

① Request Intake

The client application sends the request data + agent parameters to the Spring Boot API server via REST API over HTTPS. This includes the data to be analyzed and the target agent identifier.

② Prompt Generation

The engine queries the Prompt Registry to retrieve the latest template version for the requested agent. It then loads the Template (.md), collects the incoming request data, and performs Placeholder substitution — injecting live request data into the CONTEXT section of the template — to produce the final Assembled Prompt.

③ LLM Call

The fully assembled prompt is sent to the AI provider via multi-provider routing. The platform supports runtime switching between Gemini and GPT-4 without changing agent logic.

④ Schema Validation

The AI response is automatically validated against the predefined Output Schema (JSON Schema draft-07). Validation checks three things: required field enforcement, type validation (string, array, object), and component reuse via $ref / $defs. If validation fails, the output is flagged and logged — it never reaches the client in a broken state.

⑤ Result Delivery

The validated, schema-compliant final data is mapped to a DTO and sent back to the client application. Results are persisted to the domain-specific database before delivery when applicable.

Architecture

3-Layer Artifact Structure

registry.yaml

Registry

The agent directory. Maps each agent to its Template and Schema files, and manages versioning. A single entry per agent — enabling CI/CD-based prompt management with rollback at any time.

id name version template schema

agent_name.md

Template

The agent's behavioral specification. Written in a strict 5-section structure: MISSION → CONTEXT → RULES → Analysis Guidelines → DATA & FORMAT. The CONTEXT section contains Placeholders that receive live data at runtime.

MISSION CONTEXT RULES GUIDELINES DATA/FORMAT

agent_name_schema.json

Output Schema

The contract for AI output structure. Uses JSON Schema draft-07 with $ref / $defs for component-based type reuse. Enforces required fields, type validation, and structural consistency — guaranteeing the AI never returns a broken format.

required type $ref $defs draft-07

Prompt Engine

Template 5-Section Structure

MISSION

Purpose Definition

Defines the agent's purpose and the scope of analysis.

CONTEXT

Data Reference

Contains Placeholders replaced with live data at runtime.

RULES

Analysis Rules

Hard constraints — what the agent must always or never do.

GUIDELINES

Judgment Criteria

Domain-specific judgment criteria and thresholds.

DATA/FORMAT

Output Structure

Specifies the exact JSON output structure.

Why This Structure?

Every agent uses this exact 5-section format — no exceptions. This uniformity is what guarantees consistent output quality across all domains. A diabetes analysis agent and a geological analysis agent have completely different content, but structurally identical templates.

Placeholder Substitution

The engine reads the template, finds all Placeholder markers in the CONTEXT section, and replaces them with the actual request data. The result is the Assembled Prompt — a complete, data-rich instruction ready for the LLM.

Output Validation

Schema Validation

AI 출력 정확도·일관성을 100% 구조적으로 보장하는 게이트

LLMs produce free-form text by nature — but applications need exact JSON structures. Schema Validation is the gate that automatically closes this gap 100% of the time.

Required Field Enforcement

Every field marked required in the schema must be present in the AI output. Missing fields cause an immediate validation failure — the output is never silently incomplete.

Type Validation

Each field's data type is strictly enforced — string, array, object, number, boolean. A field that should be an array cannot arrive as a string without triggering validation failure.

$ref / $defs Component Reuse

Repeated type structures are defined once in $defs and referenced via $ref. For example, a TimelineItem type shared across blood glucose records, lab results, and medication history.

Observability · XAI

Live Broadcast Mode

파이프라인 전 단계 실시간 시각화 — Explainable AI 구현

When enabled, every stage of the pipeline streams in real time. Each step can be expanded to reveal the actual artifact used — the Registry entry, the injected Input Data, the full Template, the final Assembled Prompt, and the Output Schema used for validation. This makes AI behavior completely transparent and auditable.

Post-Deploy Prompt Verification

After deploying a new template, confirm the actual data-merged result in real time before it reaches production users.

AI Response Quality Audit

Compare the Assembled Prompt side-by-side with the AI response to evaluate output quality and prompt effectiveness.

Schema Mismatch Debugging

When validation fails, instantly see the exact diff between the Output Schema and the actual AI response to pinpoint the issue.

New Agent Development

Preview the data-injected prompt result during template authoring — validate the injection logic before the agent goes live.

Deployment

Agent Deployment Map

현재 운영 중인 에이전트 현황 — 4개 프로젝트 · 8개 에이전트

Regardless of domain, every agent uses the identical Registry → Template → Schema onboarding structure. Adding a new domain requires only creating three files — no changes to the platform itself. This is the architectural principle that makes the platform reusable across Medical, Geo, and Water domains with zero modifications to the pipeline engine.

Results

Expected Impact

AI Output Reliability

Schema Validation ensures AI output always maintains a predefined structure. Hallucination-induced irregular outputs are structurally prevented, guaranteeing accuracy and consistency across every agent invocation.

Fast Domain Onboarding

The Registry → Template → Schema structure is fully reusable. New projects launch on the same battle-tested pipeline — no matter how different the domain, the infrastructure is already in place.

Complete Audit Trail

Every pipeline stage is logged. Through XAI (Live Broadcast Mode), it is always possible to verify exactly how the AI made its decisions — providing the auditability foundation that enterprise AI systems require.