diff --git a/.coderabbit.yaml b/.coderabbit.yaml
index 00f37c1..d9de751 100644
--- a/.coderabbit.yaml
+++ b/.coderabbit.yaml
@@ -1,12 +1,15 @@
 # yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json
 language: "ko-KR"
 reviews:
-  profile: "assertive" # 깐깐하게 로직 검토
-  request_changes_workflow: false # AI가 승인을 막지 않도록 설정
+  profile: "assertive"
+  request_changes_workflow: false
   high_level_summary: true
   review_status: true
   review_details: false
-  poem: false # 불필요한 기능 제거
+  poem: false
+  pre_merge_checks:
+    docstrings:
+      mode: "off"
   auto_review:
     enabled: true
     drafts: false
@@ -14,4 +17,4 @@ reviews:
       - "develop"
       - "main"
 chat:
-  auto_reply: true
\ No newline at end of file
+  auto_reply: true
diff --git a/.github/PULL_REQUEST_TEMPLATE/release.md b/.github/PULL_REQUEST_TEMPLATE/release.md
new file mode 100644
index 0000000..b8cff89
--- /dev/null
+++ b/.github/PULL_REQUEST_TEMPLATE/release.md
@@ -0,0 +1,31 @@
+## 📌 작업 요약
+
+- 요약:
+  - develop 브랜치의 누적 변경사항을 main으로 릴리즈 배포
+- 관련 이슈: closes #
+
+## 🌿 브랜치 정보
+
+- **Source**: `develop` (기본)
+- **Target**: `main` (릴리즈)
+
+## ✅ 체크리스트
+
+- [ ] 브랜치 컨벤션 준수 (`feat/refac/hotfix/chore/design/bugfix`)
+- [ ] 커밋 컨벤션 준수 (`feat/fix/refactor/docs/style/chore`)
+- [ ] self-review 완료
+- [ ] 테스트 및 로컬 실행 확인 완료
+
+## 🧪 테스트 결과
+
+- GitHub Actions `Deploy AI to EC2` 실행 확인 (`workflow_dispatch`, ref: `main`)
+  - 결과:
+  - 스크린샷: ![ssm-send-step]()
+
+- 원격 배포 순서/재기동 확인
+  - 결과:
+  - 스크린샷: ![ssm-order]()
+
+- 배포 후 컨테이너 상태 확인
+  - 결과:
+  - 스크린샷: ![compose-ps]()
diff --git a/.github/pull_request_template.md b/.github/pull_request_template.md
index 1939f6d..2ce0da1 100644
--- a/.github/pull_request_template.md
+++ b/.github/pull_request_template.md
@@ -1,24 +1,20 @@
 ## 📌 작업 요약
-- 요약: 
+
+- 요약:
 - 관련 이슈: closes #
 
 ## 🌿 브랜치 정보
+
 - **Source**: `feat/#이슈번호-기능명`
 - **Target**: `develop` (기본) / `main` (릴리즈, 핫픽스)
 
-## 🧩 변경 타입
-- [ ] feat: 새로운 기능 추가
-- [ ] fix: 버그 수정
-- [ ] refactor: 코드 리팩토링
-- [ ] docs: 문서 수정
-- [ ] style: 코드 포맷팅, 세미콜론 누락 등
-- [ ] chore: 빌드 업무, 패키지 매니저 설정 등
-
 ## ✅ 체크리스트
+
 - [ ] 브랜치 컨벤션 준수 (`feat/refac/hotfix/chore/design/bugfix`)
 - [ ] 커밋 컨벤션 준수 (`feat/fix/refactor/docs/style/chore`)
 - [ ] self-review 완료
 - [ ] 테스트 및 로컬 실행 확인 완료
 
 ## 🧪 테스트 결과
-- (테스트 코드 실행 결과 스크린샷이나 로그, 또는 테스트 방법)
\ No newline at end of file
+
+- (테스트 코드 실행 결과 스크린샷이나 로그, 또는 테스트 방법)
diff --git a/.github/workflows/deploy-ai-ec2.yml b/.github/workflows/deploy-ai-ec2.yml
new file mode 100644
index 0000000..83c462c
--- /dev/null
+++ b/.github/workflows/deploy-ai-ec2.yml
@@ -0,0 +1,203 @@
+name: Deploy AI to EC2
+
+"on":
+  workflow_run:
+    workflows: ["AI Docker CI"]
+    types: [completed]
+    branches: ["main"]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  id-token: write
+
+env:
+  FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true
+
+jobs:
+  deploy:
+    if: github.event_name == 'workflow_dispatch' || github.event.workflow_run.conclusion == 'success'
+    runs-on: ubuntu-latest
+    concurrency:
+      group: deploy-ai-ec2-ssm
+      cancel-in-progress: false
+
+    steps:
+      - name: Resolve deploy path
+        id: deploy-path
+        env:
+          EC2_DEPLOY_PATH_SECRET: ${{ secrets.EC2_DEPLOY_PATH }}
+        run: |
+          VALUE="${EC2_DEPLOY_PATH_SECRET:-}"
+          VALUE="${VALUE%$'\r'}"
+          VALUE="$(printf '%s' "$VALUE" | sed 's/[[:space:]]*$//')"
+          if [ -n "$VALUE" ]; then
+            echo "value=$VALUE" >> "$GITHUB_OUTPUT"
+          else
+            echo "value=/home/ubuntu/GACHI-BE/deploy" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Resolve AWS region
+        id: aws-region
+        env:
+          AWS_REGION_SECRET: ${{ secrets.AWS_REGION }}
+        run: |
+          VALUE="${AWS_REGION_SECRET:-}"
+          VALUE="${VALUE%$'\r'}"
+          VALUE="$(printf '%s' "$VALUE" | sed 's/[[:space:]]*$//')"
+          if [ -n "$VALUE" ]; then
+            echo "value=$VALUE" >> "$GITHUB_OUTPUT"
+          else
+            echo "value=ap-northeast-2" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Validate deploy inputs
+        id: auth-check
+        env:
+          EC2_INSTANCE_ID: ${{ secrets.EC2_INSTANCE_ID }}
+          AWS_OIDC_ROLE_ARN: ${{ secrets.AWS_OIDC_ROLE_ARN }}
+          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+          AWS_SESSION_TOKEN: ${{ secrets.AWS_SESSION_TOKEN }}
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_USERNAME }}
+        run: |
+          set -Eeuo pipefail
+          if [ -z "${EC2_INSTANCE_ID:-}" ]; then
+            echo "::error::EC2_INSTANCE_ID secret is required."
+            exit 1
+          fi
+          if [ -z "${DOCKERHUB_USERNAME:-}" ]; then
+            echo "::error::DOCKERHUB_USERNAME secret is required."
+            exit 1
+          fi
+          if [ -n "${AWS_OIDC_ROLE_ARN:-}" ]; then
+            echo "auth_mode=oidc" >> "$GITHUB_OUTPUT"
+            echo "has_session_token=false" >> "$GITHUB_OUTPUT"
+          else
+            if [ -z "${AWS_ACCESS_KEY_ID:-}" ] || [ -z "${AWS_SECRET_ACCESS_KEY:-}" ]; then
+              echo "::error::Set AWS_OIDC_ROLE_ARN or AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY."
+              exit 1
+            fi
+            echo "auth_mode=access_key" >> "$GITHUB_OUTPUT"
+            if [ -n "${AWS_SESSION_TOKEN:-}" ]; then
+              echo "has_session_token=true" >> "$GITHUB_OUTPUT"
+            else
+              echo "has_session_token=false" >> "$GITHUB_OUTPUT"
+            fi
+          fi
+
+      - name: Configure AWS credentials with OIDC
+        if: ${{ steps.auth-check.outputs.auth_mode == 'oidc' }}
+        uses: aws-actions/configure-aws-credentials@v6
+        with:
+          aws-region: ${{ steps.aws-region.outputs.value }}
+          role-to-assume: ${{ secrets.AWS_OIDC_ROLE_ARN }}
+          role-session-name: gachi-ai-deploy-ssm
+
+      - name: Configure AWS credentials with access key
+        if: ${{ steps.auth-check.outputs.auth_mode == 'access_key' && steps.auth-check.outputs.has_session_token != 'true' }}
+        uses: aws-actions/configure-aws-credentials@v6
+        with:
+          aws-region: ${{ steps.aws-region.outputs.value }}
+          aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
+          aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+
+      - name: Configure AWS credentials with access key session
+        if: ${{ steps.auth-check.outputs.auth_mode == 'access_key' && steps.auth-check.outputs.has_session_token == 'true' }}
+        uses: aws-actions/configure-aws-credentials@v6
+        with:
+          aws-region: ${{ steps.aws-region.outputs.value }}
+          aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
+          aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+          aws-session-token: ${{ secrets.AWS_SESSION_TOKEN }}
+
+      - name: Send AI deploy command via SSM
+        id: ssm-send
+        env:
+          AWS_REGION: ${{ steps.aws-region.outputs.value }}
+          EC2_INSTANCE_ID: ${{ secrets.EC2_INSTANCE_ID }}
+          EC2_DEPLOY_PATH: ${{ steps.deploy-path.outputs.value }}
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_USERNAME }}
+        run: |
+          set -Eeuo pipefail
+          COMMANDS="$(jq -cn \
+            --arg deployPath "$EC2_DEPLOY_PATH" \
+            --arg image "$DOCKERHUB_USERNAME/gachi-ai:latest" \
+            '{
+              commands: [[
+                "set -Eeuo pipefail",
+                "cd " + ($deployPath | @sh),
+                "test -f .env",
+                "if grep -q ^AI_IMAGE= .env; then sed -i " + ("s|^AI_IMAGE=.*|AI_IMAGE=" + $image + "|" | @sh) + " .env; else echo " + ("AI_IMAGE=" + $image | @sh) + " >> .env; fi",
+                "docker compose --env-file .env pull ai",
+                "docker compose --env-file .env up -d --remove-orphans --force-recreate ai",
+                "AI_CID=$(docker compose --env-file .env ps -q ai)",
+                "test -n \"$AI_CID\"",
+                "for i in $(seq 1 24); do AI_HEALTH=$(docker inspect --format '\''{{if .State.Health}}{{.State.Health.Status}}{{else}}{{.State.Status}}{{end}}'\'' \"$AI_CID\" 2>/dev/null || echo missing); echo \"ai health: $AI_HEALTH\"; [ \"$AI_HEALTH\" = healthy ] && break; sleep 5; done",
+                "test \"$AI_HEALTH\" = healthy",
+                "docker compose --env-file .env up -d --remove-orphans --force-recreate --no-deps nginx",
+                "docker compose --env-file .env ps",
+                "docker compose --env-file .env logs --tail=80 ai nginx || true"
+              ] | join(" && ")],
+              executionTimeout: ["900"]
+            }')"
+          COMMAND_ID="$(aws ssm send-command \
+            --region "$AWS_REGION" \
+            --document-name "AWS-RunShellScript" \
+            --instance-ids "$EC2_INSTANCE_ID" \
+            --comment "deploy gachi-ai (run=${GITHUB_RUN_ID})" \
+            --max-concurrency "1" \
+            --max-errors "0" \
+            --parameters "$COMMANDS" \
+            --query "Command.CommandId" \
+            --output text)"
+          echo "command_id=$COMMAND_ID" >> "$GITHUB_OUTPUT"
+
+      - name: Wait for SSM command completion
+        env:
+          AWS_REGION: ${{ steps.aws-region.outputs.value }}
+          EC2_INSTANCE_ID: ${{ secrets.EC2_INSTANCE_ID }}
+          COMMAND_ID: ${{ steps.ssm-send.outputs.command_id }}
+        run: |
+          set -Eeuo pipefail
+          for attempt in $(seq 1 90); do
+            STATUS="$(aws ssm get-command-invocation \
+              --region "$AWS_REGION" \
+              --command-id "$COMMAND_ID" \
+              --instance-id "$EC2_INSTANCE_ID" \
+              --query "Status" \
+              --output text 2>/dev/null || true)"
+            echo "[$attempt/90] status: ${STATUS:-not-ready}"
+            case "$STATUS" in
+              Success|Failed|TimedOut|Cancelled|Cancelling)
+                break
+                ;;
+              *)
+                sleep 10
+                ;;
+            esac
+          done
+
+          INVOCATION_JSON="$(aws ssm get-command-invocation \
+            --region "$AWS_REGION" \
+            --command-id "$COMMAND_ID" \
+            --instance-id "$EC2_INSTANCE_ID" \
+            --output json)"
+          FINAL_STATUS="$(printf '%s' "$INVOCATION_JSON" | jq -r '.Status // "Unknown"')"
+          STDOUT_CONTENT="$(printf '%s' "$INVOCATION_JSON" | jq -r '.StandardOutputContent // ""')"
+          STDERR_CONTENT="$(printf '%s' "$INVOCATION_JSON" | jq -r '.StandardErrorContent // ""')"
+
+          echo "::group::SSM stdout"
+          printf '%s\n' "$STDOUT_CONTENT"
+          echo "::endgroup::"
+
+          if [ -n "$STDERR_CONTENT" ]; then
+            echo "::group::SSM stderr"
+            printf '%s\n' "$STDERR_CONTENT"
+            echo "::endgroup::"
+          fi
+
+          if [ "$FINAL_STATUS" != "Success" ]; then
+            echo "::error::SSM command failed. status=$FINAL_STATUS"
+            exit 1
+          fi
diff --git a/.github/workflows/docker-ai.yml b/.github/workflows/docker-ai.yml
index f30c075..b139edb 100644
--- a/.github/workflows/docker-ai.yml
+++ b/.github/workflows/docker-ai.yml
@@ -1,8 +1,10 @@
 name: AI Docker CI
 
-on:
+"on":
+  pull_request:
+    branches: ["develop", "main"]
   push:
-    branches: [ "develop", "main" ]
+    branches: ["develop", "main"]
   workflow_dispatch:
 
 permissions:
@@ -12,7 +14,7 @@ env:
   FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true
 
 jobs:
-  build-and-push:
+  build:
     runs-on: ubuntu-latest
 
     steps:
@@ -24,13 +26,16 @@ jobs:
         with:
           python-version: "3.11"
 
-      - name: Install deps
-        run: pip install -r requirements.txt
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r requirements.txt
 
-      - name: Syntax check
-        run: python -m py_compile app/main.py
+      - name: Compile app
+        run: python -m compileall app
 
       - name: Login Docker Hub
+        if: (github.event_name == 'push' || github.event_name == 'workflow_dispatch') && (github.ref_name == 'main' || github.ref_name == 'develop')
         uses: docker/login-action@v4
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
@@ -39,18 +44,34 @@ jobs:
       - name: Set image tags
         id: tags
         run: |
-          IMAGE="${{ secrets.DOCKERHUB_USERNAME }}/gachi-ai"
           SHORT_SHA="${GITHUB_SHA::7}"
-          if [ "${GITHUB_REF_NAME}" = "main" ]; then
-            echo "tags=${IMAGE}:latest,${IMAGE}:sha-${SHORT_SHA}" >> $GITHUB_OUTPUT
+          if [ "${GITHUB_EVENT_NAME}" = "pull_request" ]; then
+            echo "tags=gachi-ai:pr-${{ github.event.pull_request.number }}" >> "$GITHUB_OUTPUT"
+          elif [ "${GITHUB_REF_NAME}" = "main" ]; then
+            IMAGE="${{ secrets.DOCKERHUB_USERNAME }}/gachi-ai"
+            echo "tags=${IMAGE}:latest,${IMAGE}:sha-${SHORT_SHA}" >> "$GITHUB_OUTPUT"
+          elif [ "${GITHUB_REF_NAME}" = "develop" ]; then
+            IMAGE="${{ secrets.DOCKERHUB_USERNAME }}/gachi-ai"
+            echo "tags=${IMAGE}:develop,${IMAGE}:sha-${SHORT_SHA}" >> "$GITHUB_OUTPUT"
           else
-            echo "tags=${IMAGE}:develop,${IMAGE}:sha-${SHORT_SHA}" >> $GITHUB_OUTPUT
+            SAFE_REF="$(echo "${GITHUB_REF_NAME}" | tr '/' '-')"
+            echo "tags=gachi-ai:${SAFE_REF},gachi-ai:sha-${SHORT_SHA}" >> "$GITHUB_OUTPUT"
           fi
 
-      - name: Build and Push
+      - name: Build image without push
+        if: github.event_name == 'pull_request' || (github.event_name == 'workflow_dispatch' && github.ref_name != 'main' && github.ref_name != 'develop')
+        uses: docker/build-push-action@v7
+        with:
+          context: .
+          push: false
+          tags: ${{ steps.tags.outputs.tags }}
+          platforms: linux/amd64
+
+      - name: Build and push image
+        if: (github.event_name == 'push' || github.event_name == 'workflow_dispatch') && (github.ref_name == 'main' || github.ref_name == 'develop')
         uses: docker/build-push-action@v7
         with:
           context: .
           push: true
           tags: ${{ steps.tags.outputs.tags }}
-          platforms: linux/amd64
\ No newline at end of file
+          platforms: linux/amd64
diff --git a/.github/workflows/quality-ai.yml b/.github/workflows/quality-ai.yml
index d22f745..842aeec 100644
--- a/.github/workflows/quality-ai.yml
+++ b/.github/workflows/quality-ai.yml
@@ -1,10 +1,10 @@
 name: AI Quality
 
-on:
+"on":
   pull_request:
-    branches: [ "develop", "main" ]
+    branches: ["develop", "main"]
   push:
-    branches: [ "develop", "main" ]
+    branches: ["develop", "main"]
 
 permissions:
   contents: read
@@ -13,8 +13,9 @@ env:
   FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true
 
 jobs:
-  ruff:
+  test:
     runs-on: ubuntu-latest
+
     steps:
       - name: Checkout
         uses: actions/checkout@v6
@@ -24,11 +25,25 @@ jobs:
         with:
           python-version: "3.11"
 
-      - name: Install Ruff
-        run: pip install ruff
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r requirements.txt
+          pip install ruff pytest
 
       - name: Ruff check
         run: ruff check .
 
       - name: Ruff format check
         run: ruff format --check .
+
+      - name: Compile app
+        run: python -m compileall app
+
+      - name: Test if present
+        run: |
+          if [ -d tests ]; then
+            pytest
+          else
+            echo "tests 디렉터리가 없어 pytest를 건너뜁니다."
+          fi
diff --git a/README.md b/README.md
index 9f97f63..ee0601a 100644
--- a/README.md
+++ b/README.md
@@ -1,18 +1,45 @@
 # GACHI-AI
 
-GACHI 프로젝트 AI 서버(FastAPI) 레포지토리입니다.
+GACHI 프로젝트의 AI 서버입니다. 백엔드와 분리된 FastAPI 애플리케이션으로 운영하며, 기존 EC2 `docker-compose`의 `ai` 서비스로 배포합니다.
 
-## 문서
-- `docs/env.md`: 환경 변수 가이드
-- `docs/deploy.md`: 이미지 태그/배포 가이드
+## 역할
+
+- 가정통신문 원문과 날짜 후보를 기반으로 일정, 마감, 체크리스트, 알림 항목을 추출합니다.
+- OpenAI API 호출 전에도 검증할 수 있도록 비용 없는 rule-based baseline을 제공합니다.
+- 실제 LLM에 전달할 prompt-preview API를 제공합니다.
+
+## 로컬 실행
+
+```bash
+python -m venv .venv
+source .venv/bin/activate
+pip install -r requirements.txt
+uvicorn app.main:app --reload --host 0.0.0.0 --port 8000
+```
+
+Windows PowerShell에서는 다음처럼 가상환경을 활성화합니다.
+
+```powershell
+.\.venv\Scripts\Activate.ps1
+```
+
+## 주요 엔드포인트
+
+- `GET /ai/health`: 헬스체크
+- `GET /ai/docs`: Swagger UI
+- `POST /ai/newsletters/extract-items`: 날짜 후보 기반 baseline 추출
+- `POST /ai/newsletters/prompt-preview`: LLM 입력용 prompt와 response schema 미리보기
+
+## 작업 규칙
 
-## 협업 규칙
 - 기본 브랜치: `develop`
-- 브랜치 전략: `feat/xx`, `refac/xx`, `hotfix/xx`, `chore/xx`, `design/xx`, `bugfix/xx`
+- 브랜치 예시: `feat/#1-feature-name`, `chore/#1-ci-cd-setup`
 - 커밋 타입: `feat`, `fix`, `refactor`, `docs`, `style`, `chore`
-- `main`, `develop` 직접 push 금지, PR 승인 후 머지
-- CI 체크(`build-and-push`) 통과 후 머지
+- `main`, `develop` 직접 커밋은 피하고 PR로 병합합니다.
+
+## 문서
 
-## 배포 태그 규칙
-- `develop` push: `<dockerhub-id>/gachi-ai:develop`
-- `main` push: `<dockerhub-id>/gachi-ai:latest`
+- `docs/env.md`: 환경변수
+- `docs/deploy.md`: Docker image와 EC2 배포 방식
+- `docs/newsletter-extraction.md`: 가정통신문 추출 API와 프롬프트 흐름
+- `docs/newsletter-labeling-guide.md`: 정답 데이터 라벨링 기준
diff --git a/app/main.py b/app/main.py
index ecadd8e..e535acd 100644
--- a/app/main.py
+++ b/app/main.py
@@ -1,5 +1,6 @@
 from fastapi import FastAPI
-from pydantic import BaseModel
+
+from app.routers import health, newsletters
 
 app = FastAPI(
     title="GACHI-AI",
@@ -9,21 +10,5 @@
     openapi_url="/ai/openapi.json",
 )
 
-
-class EchoRequest(BaseModel):
-    text: str
-
-
-@app.get("/ai/health")
-def health() -> dict:
-    return {"status": "ok"}
-
-
-@app.get("/ai/ping")
-def ping() -> dict:
-    return {"message": "pong"}
-
-
-@app.post("/ai/echo")
-def echo(req: EchoRequest) -> dict:
-    return {"text": req.text}
+app.include_router(health.router)
+app.include_router(newsletters.router)
diff --git a/app/routers/health.py b/app/routers/health.py
new file mode 100644
index 0000000..25bdb9b
--- /dev/null
+++ b/app/routers/health.py
@@ -0,0 +1,13 @@
+from fastapi import APIRouter
+
+router = APIRouter(prefix="/ai", tags=["health"])
+
+
+@router.get("/health")
+def health() -> dict[str, str]:
+    return {"status": "ok"}
+
+
+@router.get("/ping")
+def ping() -> dict[str, str]:
+    return {"message": "pong"}
diff --git a/app/routers/newsletters.py b/app/routers/newsletters.py
new file mode 100644
index 0000000..6c34554
--- /dev/null
+++ b/app/routers/newsletters.py
@@ -0,0 +1,22 @@
+from fastapi import APIRouter
+
+from app.schemas import (
+    NewsletterExtractionRequest,
+    NewsletterExtractionResponse,
+    PromptPreviewResponse,
+)
+from app.services.newsletter_extractor import extract_newsletter_items
+from app.services.newsletter_prompt import EXTRACTION_RESPONSE_SCHEMA, build_prompt_messages
+
+router = APIRouter(prefix="/ai/newsletters", tags=["newsletters"])
+
+
+@router.post("/extract-items", response_model=NewsletterExtractionResponse)
+def extract_items(req: NewsletterExtractionRequest) -> NewsletterExtractionResponse:
+    return extract_newsletter_items(req)
+
+
+@router.post("/prompt-preview", response_model=PromptPreviewResponse)
+def prompt_preview(req: NewsletterExtractionRequest) -> PromptPreviewResponse:
+    messages = build_prompt_messages(req)
+    return PromptPreviewResponse(messages=messages, responseSchema=EXTRACTION_RESPONSE_SCHEMA)
diff --git a/app/schemas.py b/app/schemas.py
new file mode 100644
index 0000000..867c90c
--- /dev/null
+++ b/app/schemas.py
@@ -0,0 +1,89 @@
+from datetime import date
+from enum import StrEnum
+from typing import Any
+
+from pydantic import BaseModel, ConfigDict, Field, model_validator
+
+
+class ExtractedItemType(StrEnum):
+    SCHEDULE = "schedule"
+    DEADLINE = "deadline"
+    CHECKLIST = "checklist"
+    REMINDER = "reminder"
+
+
+class DateStatus(StrEnum):
+    CONFIRMED = "confirmed"
+    AMBIGUOUS = "ambiguous"
+    MISSING = "missing"
+
+
+class DateCandidate(BaseModel):
+    model_config = ConfigDict(populate_by_name=True)
+
+    candidate_id: str | None = Field(default=None, alias="candidateId")
+    original_text: str = Field(alias="originalText")
+    normalized_date: date = Field(alias="normalizedDate")
+    start_offset: int = Field(alias="startOffset", ge=0)
+    end_offset: int = Field(alias="endOffset", ge=0)
+    extraction_type: str | None = Field(default=None, alias="extractionType")
+
+    @model_validator(mode="after")
+    def validate_offsets(self) -> "DateCandidate":
+        if self.end_offset < self.start_offset:
+            raise ValueError("endOffset must be greater than or equal to startOffset")
+        return self
+
+
+class NewsletterExtractionRequest(BaseModel):
+    model_config = ConfigDict(populate_by_name=True)
+
+    original_text: str = Field(alias="originalText")
+    translated_text: str | None = Field(default=None, alias="translatedText")
+    language: str = "KO"
+    reference_date: date | None = Field(default=None, alias="referenceDate")
+    timezone: str = "Asia/Seoul"
+    date_candidates: list[DateCandidate] = Field(default_factory=list, alias="dateCandidates")
+
+
+class SelectedDateCandidate(BaseModel):
+    model_config = ConfigDict(populate_by_name=True)
+
+    index: int
+    candidate_id: str | None = Field(default=None, alias="candidateId")
+    original_text: str = Field(alias="originalText")
+    normalized_date: date = Field(alias="normalizedDate")
+
+
+class ExtractedItem(BaseModel):
+    model_config = ConfigDict(populate_by_name=True)
+
+    type: ExtractedItemType
+    title: str
+    selected_date_candidate: SelectedDateCandidate | None = Field(
+        default=None, alias="selectedDateCandidate"
+    )
+    date_status: DateStatus = Field(alias="dateStatus")
+    datetime: str | None = None
+    timezone: str
+    evidence_text: str = Field(alias="evidenceText")
+    confidence: float = Field(ge=0.0, le=1.0)
+    needs_user_confirmation: bool = Field(alias="needsUserConfirmation")
+    confirmation_question: str | None = Field(default=None, alias="confirmationQuestion")
+
+
+class NewsletterExtractionResponse(BaseModel):
+    items: list[ExtractedItem]
+    meta: dict[str, Any] = Field(default_factory=dict)
+
+
+class PromptMessage(BaseModel):
+    role: str
+    content: str
+
+
+class PromptPreviewResponse(BaseModel):
+    model_config = ConfigDict(populate_by_name=True)
+
+    messages: list[PromptMessage]
+    response_schema: dict[str, Any] = Field(alias="responseSchema")
diff --git a/app/services/newsletter_extractor.py b/app/services/newsletter_extractor.py
new file mode 100644
index 0000000..7899eb4
--- /dev/null
+++ b/app/services/newsletter_extractor.py
@@ -0,0 +1,195 @@
+import re
+from collections.abc import Iterable
+
+from app.schemas import (
+    DateCandidate,
+    DateStatus,
+    ExtractedItem,
+    ExtractedItemType,
+    NewsletterExtractionRequest,
+    NewsletterExtractionResponse,
+    SelectedDateCandidate,
+)
+
+DEADLINE_KEYWORDS = (
+    "마감",
+    "까지",
+    "제출",
+    "신청",
+    "접수",
+    "납부",
+    "등록",
+    "동의서",
+    "회신",
+)
+SCHEDULE_KEYWORDS = (
+    "일정",
+    "행사",
+    "체험",
+    "상담",
+    "설명회",
+    "교육",
+    "운영",
+    "참여",
+    "개최",
+    "방문",
+    "학습",
+)
+CHECKLIST_KEYWORDS = (
+    "준비물",
+    "지참",
+    "가져",
+    "챙겨",
+    "확인",
+    "작성",
+    "서명",
+    "제출",
+)
+
+
+def extract_newsletter_items(
+    request: NewsletterExtractionRequest,
+) -> NewsletterExtractionResponse:
+    text = request.original_text or ""
+    items = _extract_candidate_backed_items(text, request)
+    items.extend(_extract_missing_date_checklists(text, request))
+    return NewsletterExtractionResponse(
+        items=_dedupe_items(items),
+        meta={
+            "mode": "rule_based_baseline",
+            "dateCandidateCount": len(request.date_candidates),
+            "requiresLLMReview": True,
+        },
+    )
+
+
+def _extract_candidate_backed_items(
+    text: str,
+    request: NewsletterExtractionRequest,
+) -> list[ExtractedItem]:
+    items = []
+    for index, candidate in enumerate(request.date_candidates):
+        evidence = _evidence_window(text, candidate)
+        item_type = _classify_item_type(evidence)
+        selected = SelectedDateCandidate(
+            index=index,
+            candidateId=candidate.candidate_id,
+            originalText=candidate.original_text,
+            normalizedDate=candidate.normalized_date,
+        )
+        items.append(
+            ExtractedItem(
+                type=item_type,
+                title=_build_title(evidence, item_type),
+                selectedDateCandidate=selected,
+                dateStatus=DateStatus.CONFIRMED,
+                datetime=candidate.normalized_date.isoformat(),
+                timezone=request.timezone,
+                evidenceText=evidence,
+                confidence=_confidence_for(item_type),
+                needsUserConfirmation=False,
+                confirmationQuestion=None,
+            )
+        )
+    return items
+
+
+def _extract_missing_date_checklists(
+    text: str,
+    request: NewsletterExtractionRequest,
+) -> list[ExtractedItem]:
+    items = []
+    for sentence in _split_sentences(text):
+        if not _contains_any(sentence, CHECKLIST_KEYWORDS):
+            continue
+        if _overlaps_any_candidate(sentence, request.date_candidates):
+            continue
+        items.append(
+            ExtractedItem(
+                type=ExtractedItemType.CHECKLIST,
+                title=_compact_title(sentence),
+                selectedDateCandidate=None,
+                dateStatus=DateStatus.MISSING,
+                datetime=None,
+                timezone=request.timezone,
+                evidenceText=sentence,
+                confidence=0.55,
+                needsUserConfirmation=True,
+                confirmationQuestion="이 항목을 체크리스트에 추가할까요?",
+            )
+        )
+    return items
+
+
+def _classify_item_type(evidence: str) -> ExtractedItemType:
+    if _contains_any(evidence, DEADLINE_KEYWORDS):
+        return ExtractedItemType.DEADLINE
+    if _contains_any(evidence, SCHEDULE_KEYWORDS):
+        return ExtractedItemType.SCHEDULE
+    if _contains_any(evidence, CHECKLIST_KEYWORDS):
+        return ExtractedItemType.CHECKLIST
+    return ExtractedItemType.REMINDER
+
+
+def _build_title(evidence: str, item_type: ExtractedItemType) -> str:
+    title = _compact_title(evidence)
+    if item_type == ExtractedItemType.DEADLINE and "마감" not in title:
+        return f"{title} 마감"
+    return title
+
+
+def _compact_title(text: str) -> str:
+    title = re.sub(r"\s+", " ", text).strip(" \n\t-:：")
+    if len(title) <= 40:
+        return title
+    return title[:39].rstrip() + "..."
+
+
+def _evidence_window(text: str, candidate: DateCandidate) -> str:
+    if not text:
+        return candidate.original_text
+
+    start = max(candidate.start_offset - 45, 0)
+    end = min(candidate.end_offset + 70, len(text))
+    window = text[start:end]
+    left_break = max(window.rfind("\n", 0, candidate.start_offset - start), 0)
+    right_break = window.find("\n", candidate.end_offset - start)
+    if right_break == -1:
+        right_break = len(window)
+    evidence = window[left_break:right_break]
+    return re.sub(r"\s+", " ", evidence).strip() or candidate.original_text
+
+
+def _split_sentences(text: str) -> Iterable[str]:
+    for part in re.split(r"[\n.!?。]+", text):
+        sentence = re.sub(r"\s+", " ", part).strip(" -:：")
+        if sentence:
+            yield sentence
+
+
+def _overlaps_any_candidate(sentence: str, candidates: list[DateCandidate]) -> bool:
+    return any(candidate.original_text in sentence for candidate in candidates)
+
+
+def _contains_any(text: str, keywords: tuple[str, ...]) -> bool:
+    return any(keyword in text for keyword in keywords)
+
+
+def _confidence_for(item_type: ExtractedItemType) -> float:
+    if item_type in (ExtractedItemType.DEADLINE, ExtractedItemType.SCHEDULE):
+        return 0.82
+    if item_type == ExtractedItemType.CHECKLIST:
+        return 0.74
+    return 0.62
+
+
+def _dedupe_items(items: list[ExtractedItem]) -> list[ExtractedItem]:
+    seen = set()
+    result = []
+    for item in items:
+        key = (item.type, item.datetime, item.evidence_text)
+        if key in seen:
+            continue
+        seen.add(key)
+        result.append(item)
+    return result
diff --git a/app/services/newsletter_prompt.py b/app/services/newsletter_prompt.py
new file mode 100644
index 0000000..f35be02
--- /dev/null
+++ b/app/services/newsletter_prompt.py
@@ -0,0 +1,111 @@
+from app.schemas import NewsletterExtractionRequest
+
+EXTRACTION_RESPONSE_SCHEMA = {
+    "type": "object",
+    "additionalProperties": False,
+    "required": ["items"],
+    "properties": {
+        "items": {
+            "type": "array",
+            "items": {
+                "type": "object",
+                "additionalProperties": False,
+                "required": [
+                    "type",
+                    "title",
+                    "selectedDateCandidate",
+                    "dateStatus",
+                    "datetime",
+                    "timezone",
+                    "evidenceText",
+                    "confidence",
+                    "needsUserConfirmation",
+                    "confirmationQuestion",
+                ],
+                "properties": {
+                    "type": {
+                        "type": "string",
+                        "enum": ["schedule", "deadline", "checklist", "reminder"],
+                    },
+                    "title": {"type": "string"},
+                    "selectedDateCandidate": {"type": ["object", "null"]},
+                    "dateStatus": {
+                        "type": "string",
+                        "enum": ["confirmed", "ambiguous", "missing"],
+                    },
+                    "datetime": {"type": ["string", "null"]},
+                    "timezone": {"type": "string"},
+                    "evidenceText": {"type": "string"},
+                    "confidence": {"type": "number", "minimum": 0, "maximum": 1},
+                    "needsUserConfirmation": {"type": "boolean"},
+                    "confirmationQuestion": {"type": ["string", "null"]},
+                },
+            },
+        }
+    },
+}
+
+
+def build_prompt_messages(request: NewsletterExtractionRequest) -> list[dict[str, str]]:
+    return [
+        {"role": "system", "content": _build_system_prompt()},
+        {"role": "user", "content": _build_user_prompt(request)},
+    ]
+
+
+def _build_system_prompt() -> str:
+    return """
+역할: 학교 가정통신문에서 캘린더, 알림, 체크리스트로 만들 항목을 추출한다.
+
+핵심 규칙:
+- 구체적인 날짜는 반드시 제공된 date candidates 중 하나만 선택할 것.
+- date candidates에 없는 날짜를 새로 만들거나 추론하지 말 것.
+- 날짜 근거가 명확할 때만 dateStatus를 "confirmed"로 설정할 것.
+- 날짜 후보가 없거나 근거가 약하면 "ambiguous" 또는 "missing"을 사용할 것.
+- evidenceText는 원문에서 짧게 가져올 것.
+- response schema에 맞는 JSON만 반환할 것.
+
+항목 분류 기준:
+- deadline: 제출, 신청, 납부, 등록, 동의, 회신, 마감 행동
+- schedule: 행사, 수업, 상담, 체험학습, 설명회, 운영일
+- checklist: 준비물, 지참물, 확인 문서, 보호자나 학생이 해야 할 행동
+- reminder: deadline이나 schedule은 아니지만 알림으로 보여줄 가치가 있는 항목
+""".strip()
+
+
+def _build_user_prompt(request: NewsletterExtractionRequest) -> str:
+    translated_text = request.translated_text.strip() if request.translated_text else ""
+    reference_date = request.reference_date.isoformat() if request.reference_date else "null"
+    sections = [
+        f"referenceDate: {reference_date}",
+        f"timezone: {request.timezone}",
+        f"language: {request.language}",
+        "",
+        "<date_candidates>",
+        _format_candidates(request),
+        "</date_candidates>",
+        "",
+        "<original_text>",
+        request.original_text.strip(),
+        "</original_text>",
+    ]
+    if translated_text:
+        sections.extend(["", "<translated_text>", translated_text, "</translated_text>"])
+    return "\n".join(sections)
+
+
+def _format_candidates(request: NewsletterExtractionRequest) -> str:
+    if not request.date_candidates:
+        return "[]"
+
+    lines = []
+    for index, candidate in enumerate(request.date_candidates):
+        lines.append(
+            f"- index: {index}, "
+            f"candidateId: {candidate.candidate_id or 'null'}, "
+            f"originalText: {candidate.original_text}, "
+            f"normalizedDate: {candidate.normalized_date.isoformat()}, "
+            f"startOffset: {candidate.start_offset}, "
+            f"endOffset: {candidate.end_offset}"
+        )
+    return "\n".join(lines)
diff --git a/docs/deploy.md b/docs/deploy.md
index db88849..9d86667 100644
--- a/docs/deploy.md
+++ b/docs/deploy.md
@@ -1,42 +1,43 @@
-# GACHI-AI Deploy Guide
+# GACHI-AI 배포 가이드
 
-## 1) 이미지 태그 규칙
-- `develop` 브랜치 push: `<DOCKERHUB_USERNAME>/gachi-ai:develop`, `sha-<7자리>`
-- `main` 브랜치 push: `<DOCKERHUB_USERNAME>/gachi-ai:latest`, `sha-<7자리>`
+## Docker image 태그
+
+- `develop` push: `<DOCKERHUB_USERNAME>/gachi-ai:develop`, `<DOCKERHUB_USERNAME>/gachi-ai:sha-xxxxxxx`
+- `main` push: `<DOCKERHUB_USERNAME>/gachi-ai:latest`, `<DOCKERHUB_USERNAME>/gachi-ai:sha-xxxxxxx`
+
+## GitHub Actions secrets
+
+Docker image build/push에 필요합니다.
 
-## 2) GitHub Actions 필수 시크릿
 - `DOCKERHUB_USERNAME`
 - `DOCKERHUB_TOKEN`
 
-## 3) EC2 반영 (BE compose에서 함께 기동)
+EC2 배포에 필요합니다.
 
-### develop 반영
-```bash
-cd ~/GACHI-BE/deploy
-sed -i 's|^AI_IMAGE=.*|AI_IMAGE=<DOCKERHUB_USERNAME>/gachi-ai:develop|' .env
-docker compose --env-file .env pull ai
-docker compose --env-file .env up -d --force-recreate ai nginx
-```
+- `EC2_INSTANCE_ID`
+- `AWS_REGION`
+- `AWS_OIDC_ROLE_ARN` 또는 `AWS_ACCESS_KEY_ID` / `AWS_SECRET_ACCESS_KEY`
+- 선택: `AWS_SESSION_TOKEN`
+- 선택: `EC2_DEPLOY_PATH` 기본값은 `/home/ubuntu/GACHI-BE/deploy`
 
-### main 반영
-```bash
-cd ~/GACHI-BE/deploy
-sed -i 's|^AI_IMAGE=.*|AI_IMAGE=<DOCKERHUB_USERNAME>/gachi-ai:latest|' .env
-docker compose --env-file .env pull ai
-docker compose --env-file .env up -d --force-recreate ai nginx
-```
+## 배포 방식
+
+AI 서버는 별도 EC2를 만들지 않고 기존 백엔드 EC2의 compose 파일에 정의된 `ai` 서비스로 배포합니다.
+
+`main` push에서 `AI Docker CI`가 성공하거나, `workflow_dispatch`로 `deploy-ai-ec2.yml`을 직접 실행하면 다음 작업을 수행합니다.
+
+1. EC2의 deploy path로 이동
+2. `.env`의 `AI_IMAGE`를 `<DOCKERHUB_USERNAME>/gachi-ai:latest`로 갱신
+3. `docker compose --env-file .env pull ai`
+4. `ai` 컨테이너 재생성
+5. `ai` health check 확인
+6. 필요 시 `nginx` 재생성
+
+## 수동 확인
 
-## 4) 상태 확인
 ```bash
+cd /home/ubuntu/GACHI-BE/deploy
 docker compose --env-file .env ps
+curl -i http://localhost:8000/ai/health
 curl -i http://localhost/ai/health
-curl -i http://localhost/ai/docs
-```
-
-## 5) 롤백
-```bash
-cd ~/GACHI-BE/deploy
-sed -i 's|^AI_IMAGE=.*|AI_IMAGE=<DOCKERHUB_USERNAME>/gachi-ai:sha-<원하는태그>|' .env
-docker compose --env-file .env pull ai
-docker compose --env-file .env up -d --force-recreate ai nginx
 ```
diff --git a/docs/env.md b/docs/env.md
index 3b19f18..8cc1b13 100644
--- a/docs/env.md
+++ b/docs/env.md
@@ -1,7 +1,13 @@
-# AI Environment Variables
+# AI 서버 환경변수
 
-## Required
-- `OPENAI_API_KEY`: LLM API key
+## 필수
 
-## Optional
-- `LOG_LEVEL`: default `INFO`
\ No newline at end of file
+- `OPENAI_API_KEY`: 추후 LLM API 호출을 붙일 때 사용할 OpenAI API key
+
+## 선택
+
+- `LOG_LEVEL`: 로그 레벨. 기본값은 `INFO`
+
+## 현재 상태
+
+현재 구현은 OpenAI API를 직접 호출하지 않습니다. `OPENAI_API_KEY`는 기존 EC2 compose 환경과 향후 LLM client 연결을 고려해 유지합니다.
diff --git a/docs/newsletter-extraction.md b/docs/newsletter-extraction.md
new file mode 100644
index 0000000..b626a5a
--- /dev/null
+++ b/docs/newsletter-extraction.md
@@ -0,0 +1,55 @@
+# 가정통신문 항목 추출 설계
+
+## 목적
+
+가정통신문 본문에서 일정, 마감, 체크리스트, 알림 항목을 추출합니다.
+
+핵심 원칙은 날짜를 AI가 새로 만들지 않게 하는 것입니다. 구체적인 날짜는 백엔드나 전처리 단계에서 만든 `dateCandidates` 중 하나만 선택해야 합니다.
+
+## 엔드포인트
+
+### `POST /ai/newsletters/extract-items`
+
+비용 없이 실행되는 rule-based baseline입니다. OpenAI API를 붙이기 전에도 스키마, 날짜 후보 매칭, 샘플 케이스를 확인할 수 있습니다.
+
+### `POST /ai/newsletters/prompt-preview`
+
+LLM에 전달할 system/user prompt와 응답 JSON schema를 생성합니다. ChatGPT Plus에서 수동 실험하거나, 추후 OpenAI API 호출에 그대로 사용할 수 있습니다.
+
+## 요청 예시
+
+```json
+{
+  "originalText": "5월 10일까지 참가 신청서를 제출해주세요.",
+  "translatedText": null,
+  "language": "KO",
+  "referenceDate": "2026-05-06",
+  "timezone": "Asia/Seoul",
+  "dateCandidates": [
+    {
+      "candidateId": "dc_1",
+      "originalText": "5월 10일",
+      "normalizedDate": "2026-05-10",
+      "startOffset": 0,
+      "endOffset": 6,
+      "extractionType": "REGEX"
+    }
+  ]
+}
+```
+
+## 추출 규칙
+
+- `confirmed`는 항목이 제공된 date candidate 중 하나를 사용할 때만 부여합니다.
+- 날짜 표현이 있지만 후보 매칭이 불확실하면 `ambiguous`로 둡니다.
+- 실행 가능한 체크리스트인데 날짜가 없으면 `missing`으로 둡니다.
+- `evidenceText`는 원문 근거를 짧게 담습니다.
+- 캘린더와 알림 생성은 `confirmed` 항목만 대상으로 삼습니다.
+
+## 작업 흐름
+
+1. 백엔드 또는 전처리 단계에서 날짜 후보를 만듭니다.
+2. AI 서버에 원문과 `dateCandidates`를 전달합니다.
+3. `extract-items`로 비용 없는 baseline 결과를 먼저 확인합니다.
+4. 부족한 케이스는 `prompt-preview` 결과를 ChatGPT Plus에 넣어 비교합니다.
+5. 충분히 안정화된 뒤 OpenAI API 호출을 AI 서버 내부에 붙입니다.
diff --git a/docs/newsletter-labeling-guide.md b/docs/newsletter-labeling-guide.md
new file mode 100644
index 0000000..b347776
--- /dev/null
+++ b/docs/newsletter-labeling-guide.md
@@ -0,0 +1,39 @@
+# 가정통신문 라벨링 기준
+
+본 문서는 사람이 직접 구축하는 정답 데이터와 AI 모델의 판단 기준을 일치시키기 위한 라벨링 가이드라인입니다.
+
+## 기본 원칙
+
+- 원문 근거가 있는 항목만 라벨링한다.
+- 날짜는 제공된 `dateCandidates` 중 하나만 선택한다.
+- 날짜 후보에 없는 날짜를 사람이 임의로 만들지 않는다.
+- 캘린더/알림 생성은 `dateStatus=confirmed` 항목만 대상으로 본다.
+
+## 항목 분류
+
+- `deadline`: 제출, 신청, 납부, 등록, 동의, 회신, 마감처럼 특정 날짜까지 해야 하는 행동
+- `schedule`: 행사, 수업, 상담, 체험학습, 설명회, 운영일처럼 실제로 진행되는 일정
+- `checklist`: 준비물, 지참물, 확인할 문서, 보호자나 학생이 해야 할 행동
+- `reminder`: deadline이나 schedule은 아니지만 알림으로 보여줄 가치가 있는 정보
+
+## 날짜 상태
+
+- `confirmed`: 원문 근거와 `dateCandidates` 중 하나가 명확히 연결된 상태
+- `ambiguous`: 날짜 표현은 있지만 후보 매칭이 불확실한 상태
+- `missing`: 행동 항목은 있지만 날짜가 없는 상태
+
+## 라벨링 예시
+
+```json
+{
+  "type": "deadline",
+  "title": "체험학습 동의서 제출",
+  "evidenceText": "5월 20일까지 체험학습 동의서를 제출해주세요.",
+  "selectedDateCandidateId": "dc_1",
+  "dateStatus": "confirmed",
+  "date": "2026-05-20",
+  "target": "parent",
+  "actionRequired": true,
+  "schoolContext": "보호자가 동의서를 확인하고 제출해야 하는 안내"
+}
+```
diff --git a/requirements.txt b/requirements.txt
index f037fb5..60d3c14 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,2 +1,3 @@
 fastapi==0.116.1
-uvicorn[standard]==0.35.0
\ No newline at end of file
+pydantic>=2.0,<3.0
+uvicorn[standard]==0.35.0