跳至主要内容

[Maeumjigi 畢業設計開發日誌 #23] 雙層 Safety Wrapper 護欄、AWS EC2 GPU 算力成本與第一學期答辯大捷

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

Cover Image

June 18, 2026, marks the triumphant completion of the 1st semester of our university capstone engineering project. From our humble beginnings as a simple Android convention scraper in March, Maeumjigi has evolved into a full-scale, clinical-grade AI mental health counseling ecosystem.

Our final sprint focused on two mission-critical objectives:

  1. The Dual-Layer Safety Wrapper: Engineering a bulletproof guardrail architecture to guarantee zero-hallucination crisis intervention.
  2. Production Cloud Serving & Final Defense: Sizing AWS EC2 GPU infrastructure budgets and delivering our final faculty presentation.

1. The Dual-Layer Safety Wrapper Architecture​

In consumer chatbots, occasional hallucinations or awkward phrasing are minor annoyances. In an AI platform dedicated to youth and military suicide prevention, a single hallucinated response or dismissive answer could be fatal.

To protect vulnerable users, we designed a Dual-Layer Safety Wrapper that wraps around our fine-tuned Qwen 3.5-9B model (Checkpoint-4400):

flowchart TD
    subgraph Inbound_Triage ["Layer 1: Pre-Inference Guardrail"]
        A[User Message Received] --> B[Regex Crisis Keyword Detector]
        B --> C[Zero-Shot Semantic Crisis Classifier]
        C --> D{Risk Severity Assessment}
    end

    subgraph Decision_Branch ["Emergency Routing"]
        D -->|Level 3: Active Suicidality / Imminent Danger| E[Emergency Protocol Triggered]
        E -->|Bypass Generative Model| F[Deterministic Crisis Intervention & 109 / 1393 Dialing]
        
        D -->|Level 1-2: Mild to Moderate Emotional Distress| G[Pass to Fine-Tuned LLM]
    end

    subgraph Inference_Core ["Model Generation"]
        G --> H[vLLM Serving Engine: Checkpoint-4400]
        H --> I[Raw Generative Token Stream]
    end

    subgraph Outbound_Verifier ["Layer 2: Post-Inference Guardrail"]
        I --> J[Toxicity & Hallucination Filter]
        J --> K[Clinical Referral Footer Injection]
        K --> L[Sanitized Real-Time SSE to Client]
    end

    F --> L

Crisis Escalation Protocol (Level 1 to 3)​

Risk ClassificationClinical IndicatorsSystem Response ActionLatency Target
Level 1 (Mild)Fatigue, loneliness, academic/work stressEmpathetic listening, active validation< 800 ms
Level 2 (Moderate)Chronic hopelessness, self-worth collapseCognitive reframing, counseling clinic recommendation< 800 ms
Level 3 (Crisis)Explicit suicidal ideation, intent, or self-harmInstant LLM bypass: Display 24/7 Lifeline (109 / 1393) modal with one-tap emergency call< 150 ms

2. Guardrail Middleware Implementation​

The Safety Wrapper operates as an asynchronous FastAPI middleware, validating payloads prior to dispatching GPU inference requests:

# Safety Wrapper core logic: safety_wrapper.py
import re
from fastapi import Request, HTTPException

CRISIS_PATTERNS = [
    r"(죽고\s*싶|자살|목숨을\s*끊|살기\s*싫|뛰어내리)",
    r"(suicide|kill\s*myself|end\s*my\s*life)"
]

class SafetyWrapper:
    def __init__(self, emergency_hotline="109"):
        self.compiled_regex = [re.compile(p, re.IGNORECASE) for p in CRISIS_PATTERNS]
        self.hotline = emergency_hotline

    def evaluate_input(self, user_text: str) -> dict:
        for pattern in self.compiled_regex:
            if pattern.search(user_text):
                return {
                    "is_crisis": True,
                    "action": "ESCALATE_IMMEDIATELY",
                    "override_response": (
                        f"당신의 고통을 깊이 공감합니다. 지금 즉시 전문가의 도움을 받으실 수 있습니다. "
                        f"24시간 자살예방 상담전화 ☎ {self.hotline}번으로 연락해 주세요. 당신은 혼자가 아닙니다."
                    )
                }
        return {"is_crisis": False, "action": "PROCEED_TO_LLM"}

3. AWS EC2 GPU Production Budget & Sizing​

To serve our validated chatbot to university students and military beta testers over the 5-month summer break, we conducted an AWS EC2 instance sizing and cost optimization study:

Instance FamilyGPU ConfigurationvCPU / RAMOn-Demand (Hourly)5-Month Estimated Cost (On-Demand)1-Year Savings Plan (Hourly)5-Month Optimized Cost
g4dn.2xlarge1x NVIDIA T4 (16GB)8 vCPU / 32GB$0.752$2,707$0.481$1,731
g5.2xlarge (Chosen)1x NVIDIA A10G (24GB)8 vCPU / 32GB$1.212$4,363$0.763$2,746 (Save 37%)
g6.2xlarge1x NVIDIA L4 (24GB)8 vCPU / 32GB$0.978$3,520$0.626$2,253

Engineering Selection: While g4dn.2xlarge offered lower cost, its 16GB VRAM could not host our 9B model alongside vLLM PagedAttention KV-caches without extreme 4-bit quantization degradation. We selected g5.2xlarge with a 1-year Savings Plan, unlocking 24GB high-speed GDDR6 VRAM, sub-50ms token latencies, and 37% budgetary savings!

Final Presentation

4. 1st Semester Capstone Retrospective & Defense Highlights​

On June 18th, our engineering team presented the completed 1st Semester Final Defense Presentation (PPTX) before the university evaluation panel:

timeline
    title Maeumjigi Semester 1 Engineering Milestones
    March 2026 : Project Kickoff : Claude Code Adoption : Convention Scraper Pivot
    April 2026 : 120K Corpus Preprocessing : Llama 3.1 JSONL Pipeline : Next.js 14 Frontend
    May 2026 : 4 Rounds of Fine-Tuning : Checkpoint-4400 Locked : FastAPI PostgreSQL Hardened
    June 2026 : Dual Safety Wrapper : AWS g5.2xlarge Sizing : 1st Semester Final Presentation Passed

The review panel commended the team for solving the over-counseling trap through empirical data mixing (2:1 ratio) and praised the clinical safety guarantees provided by the deterministic Safety Wrapper.

5. What's Next in Semester 2?​

With Phase 1 concluded, our team enters summer deployment:

  1. Provisioning our AWS EC2 g5.2xlarge instance with automated containerized health checks.
  2. Launching closed-beta testing across military welfare communities.
  3. Returning in Semester 2 with live multi-turn field data and expanded knowledge graph capabilities!