跳至主要内容

[Maeumjigi 畢業設計開發日誌 #1] 第一步與規劃: Claude Code AI 協作規則及從預約提醒向心理健康 AI 的關鍵轉型

· 閱讀時間約 4 分鐘
Jihwan Park (jaybee)
Jihwan Park (jaybee)
AI & Robotics Engineer

Cover Image

March 26, 2026, marks the historic first day of our university capstone engineering project. Today laid the structural bedrock for everything to come: adopting an autonomous AI coding assistant (Claude Code), designing team collaboration protocols, and making the pivotal architectural decision that transformed a simple reservation alert tool into Maeumjigi, a specialized AI mental health counseling platform.

1. The Autonomous AI Collaboration Framework: Claude Code & Rulesets​

At the onset of the capstone project, our engineering team decided to break away from traditional manual pair programming and integrate Claude Code as our primary autonomous coding agent.

To prevent AI hallucination, uncontrolled code sprawl, and incompatible dependency conflicts, we established two foundational rule documents: Claude.md (project domain constraints) and CLAUDE_coding_rule.md (syntactic and architectural rules).

# Excerpt from CLAUDE_coding_rule.md
1. Architecture Integrity:
   - All backend APIs must follow strict RESTful standards using FastAPI and Pydantic schemas.
   - Do not invent non-existent packages; verify imports against requirements.txt.
2. Safe Automation:
   - The AI agent has autonomous permissions within the project workspace.
   - Irreversible file deletions or git push operations require explicit verification checkpoints.
3. Documentation & Verification:
   - Every module must maintain unit tests covering standard execution and edge-case exceptions.

By defining these boundaries on Day 1, Claude Code transformed from a simple autocomplete tool into an autonomous junior engineer capable of running data pipelines, debugging database crashes, and compiling comprehensive technical documentation.

2. The Architectural Pivot: From Alarm Notifier to Mental Health AI​

Our team's initial capstone proposal was a practical Android utility: a web scraper monitoring wedding convention cancellations (MND Convention) that delivered instant mobile push notifications via Firebase Cloud Messaging (FCM).

However, during architectural review, we recognized that simple web scraping had limited academic depth and high legal/operational fragility due to anti-bot protections. Concurrently, analyzing societal data revealed a severe crisis: rising depressive symptoms and psychological isolation among young adults and military service members, contrasted with a stark shortage of licensed clinical psychologists.

We executed an immediate architectural pivot: repurposing our engineering pipeline to build a 24/7 empathetic crisis intervention chatbot fine-tuned on real Korean empathetic dialogue datasets.

flowchart TD
    subgraph Initial_Plan ["Initial Concept: Android Convention Alarm"]
        A[MND Convention Web Portal] -->|Periodic Scraper| B[Python Polling Engine]
        B -->|Status Change Detected| C[FCM Push Server]
        C -->|Instant Alert| D[Android Client App]
    end

    subgraph Architecture_Pivot ["Critical Engineering Pivot"]
        direction TB
        E[Societal Problem: Youth & Military Mental Health Crisis]
        F[Academic Rigor: Fine-Tuned Open-Source LLMs]
    end

    subgraph Final_Roadmap ["Final System: Maeumjigi Mental Health AI"]
        G[AI Hub Korean Empathetic Corpora] -->|100K Curation| H[Llama 3.1 & Qwen Fine-Tuning]
        H -->|LoRA & QLoRA| I[Dual-Layer Safety Wrapper]
        I -->|FastAPI Serving| J[Cross-Platform Client Web/App]
    end

    Initial_Plan -.->|Evaluated for Capstone Scope| Architecture_Pivot
    Architecture_Pivot ==>|System Redeveloped| Final_Roadmap

3. Engineering Feasibility Comparison​

To validate our pivot before the faculty review board, we conducted a side-by-side engineering feasibility matrix:

Evaluation DimensionInitial Concept (Convention Alarm App)Pivoted Architecture (Maeumjigi AI Platform)
Primary Tech StackAndroid (Kotlin), Python BeautifulSoup, FCMFastAPI, PyTorch, Hugging Face, Llama 3.1, React Native
Data EngineeringFragile HTML DOM parsing & cron jobs100,000+ multi-turn empathetic dialogue curation
Algorithmic DepthDeterministic string matching & diff checksLoRA fine-tuning, loss convergence, Perplexity evaluation
Safety RequirementsLow (occasional false alarm acceptable)Critical (Active Suicidality triage & zero hallucination)
Scalability & ImpactLimited to single event reservation utilityUniversal mental health support and suicide prevention triage

4. Prototype Scraper vs. Prompt Engineering Blueprint​

On Day 1, we preserved our exploratory scraper prototype while drafting the prompt engineering blueprint for the mental health model:

# Initial prototyping: Establishing structured logging and error handling
import sys
import logging

logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")

def initialize_project_environment():
    logging.info("Initializing Capstone Workspace...")
    rules = ["CLAUDE_coding_rule.md", "Claude.md"]
    for r in rules:
        logging.info(f"Enforcing active agent configuration: {r}")
    logging.info("Core system status: Ready for Phase 1 Data Pipeline.")

if __name__ == "__main__":
    initialize_project_environment()

Terminal Screenshot

5. Day 1 Key Achievements & Next Steps​

  1. Autonomous Workflow Setup: Established Claude Code coding rules (CLAUDE_coding_rule.md) to ensure zero code divergence during automated generations.
  2. Capstone Direction Finalized: Successfully approved the pivot from a trivial scraper app to a high-impact mental health LLM platform.
  3. Data Acquisition Strategy: Outlined the pipeline to source and filter over 100,000 multi-turn counseling dialogues from AI Hub.

In the next devlog, we will explore our initial Android and React Native UI scaffolding and the first attempts at data ingestion!