곁 (Gyeot) v0.1 — Project Specification
Tagline: 너의 곁에서, 너와 함께. / By your side, together with you.
Organization: 22B Labs · The 4th Path (⟨H⊕A⟩ ↦ Ω)
License: MIT
Repository: github.com/sinmb79/gyeot (planned)
Last Updated: 2026-04-04
1. Vision & Philosophy
Gyeot is a relationship-based AI companion for people who have no one to ask.
Not a chatbot. Not a search engine. Not a therapist.
A parent who asks "did you eat?", a senior colleague who says "everyone struggles at first", a neighbor who says "let me help you with that form."
Gyeot exists because the question "who do I call?" has no answer for millions of people — foster youth aging out of care, single parents raising children alone, young carers sacrificing their futures, elderly living in isolation, newcomers navigating an unfamiliar society.
The 4th Path philosophy applies directly: AI should not extract from vulnerable people. It should stand beside them. Gyeot is the bridge, never the destination. The goal is graduation — the day a user no longer needs Gyeot because they have real people around them.
2. Target Users & Their Needs
2.1 자립준비청년 (Foster Youth Aging Out)
- Scale: ~2,000/year exit care; ~10,000 cumulative within 5 years
- Core needs: Administrative life skills (registration, contracts, insurance), fraud prevention, emotional support, meal/health check-ins
- AI role: Parent (부모)
- Key stat: Mental illness rate 12.7%; life satisfaction 5.6/10 vs 6.7 national youth average
2.2 한부모가정 (Single-Parent Families)
- Scale: ~1.5M households
- Core needs: Childcare advice, education guidance, career path for children, welfare benefit navigation, emotional support
- AI role: Co-parent / Supportive relative (가족)
- Key stat: Income at 60.3% of national average; top difficulty is childcare cost burden
2.3 가족돌봄청년 — Young Carers
- Scale: 57,000–632,000 estimated
- Core needs: Stress relief, career/education info for themselves, welfare service connection, validation
- AI role: Supportive senior (선배)
- Key stat: Depression rate 61.5% (vs 8.5% non-carer youth); avg 21.6 hrs/week caregiving
2.4 고립·은둔 청년 (Isolated/Withdrawn Youth)
- Scale: 5.2% of youth population
- Core needs: Low-pressure first contact, gradual social reintegration guidance, mental health resource connection
- AI role: Patient friend (친구)
- Key stat: 6.3% needed professional counseling but couldn't access; top barrier is cost (38.6%)
2.5 독거노인 / 고독사 위험군 (Elderly Living Alone / Lonely Death Risk)
- Scale: 3,924 lonely deaths in 2024 (+7.2% YoY); 33% of adults have no one to turn to
- Core needs: Daily check-ins, medication reminders, fraud prevention, companionship, emergency connection
- AI role: Caring child/neighbor (자녀/이웃)
- Key stat: 81.7% male; 60s (32.4%) and 50s (30.5%) most vulnerable
2.6 사회초년생 (Career Starters without Support Network)
- Scale: Subset of 19–34 youth population
- Core needs: Workplace skills (Excel, reports, dress code, drinking etiquette), career advice, burnout management
- AI role: Workplace senior (직장 선배)
- Key stat: 32.2% experience burnout; top cause is career anxiety (39.1%)
2.7 다문화가정 (Multicultural Families)
- Core needs: Korean life information in native language + easy Korean, cultural guidance, child education support
- AI role: Neighbor / local guide (이웃 언니/형)
2.8 군 전역자 (Military Discharge)
- Core needs: Civilian reintegration, job prep, social skill recovery after 2-year gap
- AI role: Senior veteran (선임 선배)
3. Core Features
3.1 Relationship-Based Roles
Gyeot is NOT a generic assistant. It operates in defined relationship roles:
| Role ID | Role Name | Tone | Example Behavior |
|---|---|---|---|
parent |
부모 | Warm, caring, practical | "밥은 먹었어?", "내일 면접이지? 일찍 자" |
senior |
선배 | Encouraging, experienced | "처음엔 다 그래, 괜찮아", "엑셀 단축키 알려줄까?" |
sibling |
형제/자매 | Casual, supportive | "야 그거 사기야, 절대 보내지 마" |
neighbor |
이웃 | Friendly, helpful | "이 동네 병원은 여기가 좋아" |
sponsor |
후원자 | Respectful, empowering | "네가 해낸 거야, 대단하다" |
Users can select a role or let AI infer from onboarding conversation.
3.2 Proactive Care (AI-Initiated Contact)
Gyeot reaches out first. Real family doesn't wait to be asked.
Triggers:
- Calendar events → "내일 면접이야, 준비물 챙겼어?"
- Meal times → "점심 먹었어?" (configurable frequency)
- Welfare deadlines → "자립수당 신청 마감 3일 남았어"
- Weather → "비 온다는데 우산 챙겨"
- Inactivity → "요즘 조용하네, 괜찮아?"
- Achievement → "이번 달 저축 목표 달성했어! 잘했다"
3.3 Conversational Onboarding & Auto-Profiling
No survey forms. Natural conversation flow:
"안녕! 너를 좀 알고 싶어. 지금 혼자 살아?"
→ User responds naturally
→ Internal profile auto-generated:
{
family_status: "no_parents",
living: "alone",
age_group: "late_teens",
current_activity: "job_seeking",
risk_flags: ["social_isolation"],
activated_modules: ["parent_role", "welfare_scanner", "meal_checkin", "fraud_alert"],
tone: "casual_warm"
}
Rules:
- NEVER label the user ("You are in a vulnerable group")
- NEVER display internal profile categories
- User sees: "네 상황에 맞게 준비했어"
- User can manually adjust in Settings → "나의 상황"
- Profile updates dynamically as user's life changes
3.4 Life Integration Layer
Calendar Integration (Google/Apple/Kakao)
- Read events → context-aware encouragement
- Suggest events → "자립수당 신청일 등록해둘까?"
- Pre-event care → "면접 전날이야, 일찍 자"
Financial Awareness (Built-in simple ledger or Toss/BankSalad link)
- Spending pattern recognition → gentle advice
- "이번 달 식비가 좀 많았어, 다음 주는 집밥 어때?"
- Bill due date reminders
- NEVER judgmental, always supportive
Health Tracking
- Medication reminders (elderly, chronic illness)
- Hospital appointment awareness
- "약 먹을 시간이야"
Welfare Benefit Scanner
- User profile → auto-scan eligible benefits from 복지로/자립정보ON/정부24
- "너 이거 해당되는데 신청했어?"
- Links to official application pages
- RAG-powered with government data sources
3.5 Growth Tracking
- Milestone recognition: "6개월 전엔 전입신고 방법 물어봤는데, 이제 연말정산 혼자 하네. 많이 컸다"
- Progress visualization (optional)
- Graduation encouragement → mentor transition
3.6 Bridge to Humans
- Connect to 바람개비서포터즈 (foster youth mentors)
- Connect to 자립지원전담기관 (independence support agencies)
- Peer matching (anonymized, similar-situation users)
- Community events recommendation
4. Architecture
4.1 Overview
┌─────────────────────────────────────────────┐
│ USER DEVICE │
│ │
│ ┌─────────────────────────────────────┐ │
│ │ Gyeot App (PWA) │ │
│ │ ┌──────────┐ ┌────────────────┐ │ │
│ │ │ On-Device │ │ Local Storage │ │ │
│ │ │ LLM │ │ (encrypted) │ │ │
│ │ │ Gemma 4 │ │ - Profile │ │ │
│ │ │ E4B/E2B │ │ - Chat history│ │ │
│ │ └──────────┘ │ - Calendar │ │ │
│ │ │ - Finance │ │ │
│ │ └────────────────┘ │ │
│ └──────────────┬──────────────────────┘ │
│ │ (only when needed) │
└─────────────────┼───────────────────────────┘
│
▼
┌─────────────────────────────────────────────┐
│ GYEOT PROXY SERVER │
│ (Stateless — no user data stored) │
│ │
│ ┌──────────────┐ ┌───────────────────┐ │
│ │ Sponsor Key │ │ RAG Engine │ │
│ │ Pool │ │ (Government DB) │ │
│ │ │ │ - 복지로 API │ │
│ │ Key A ──┐ │ │ - 자립정보ON │ │
│ │ Key B ──┤ │ │ - 정부24 │ │
│ │ Key C ──┘ │ │ - Legal DB │ │
│ └──────────────┘ └───────────────────┘ │
│ │
│ ┌──────────────────────────────────────┐ │
│ │ Crisis Detection │ │
│ │ → 1393 (Suicide Prevention) │ │
│ │ → 112 (Child Protection) │ │
│ │ → 1644-6621 (Single Parent) │ │
│ │ → 119 (Emergency) │ │
│ └──────────────────────────────────────┘ │
└─────────────────────────────────────────────┘
4.2 On-Device Layer (Free Tier — Unlimited)
- Primary: Gemma 4 E4B (8GB+ RAM phones)
- Fallback: Gemma 4 E2B (6GB RAM phones)
- Runtime: ExecuTorch (production) / llama.cpp (prototype)
- Capabilities:
- Daily conversation, emotional support, basic advice
- Calendar/finance integration (all local)
- Proactive care triggers (all local)
- Native audio input/output (no separate STT/TTS needed)
- Offline operation — works without internet
- Data: ALL personal data stays on device, encrypted
4.3 Cloud Layer (Sponsor-Funded Tier)
- Trigger: User taps "후원이 필요해요" (one-time, grants 30-day access)
- Proxy server:
- Receives anonymized query from user app
- Selects available sponsor API key from pool
- Calls cloud LLM (Gemini/DeepSeek/Qwen/OpenAI — whatever key the sponsor registered)
- Returns response to user
- Logs usage count only (NO conversation content)
- Use cases: Complex welfare eligibility checks (RAG), contract/lease review, multi-step administrative guidance
- Sponsor dashboard: Usage stats only — "Your key served 127 conversations this month"
4.4 KakaoTalk Channel (Digital Divide Bridge)
- For elderly and users who cannot install apps
- Server-side processing via sponsor API keys
- Simplified interface, voice message support
- Emergency button always visible
5. Sponsorship Model
5.1 API Key Donation
Sponsors register their own LLM API keys on the Gyeot platform.
Flow:
- Sponsor visits gyeot.the4thpath.com/sponsor
- Registers API key (encrypted, stored on proxy server)
- Sets monthly budget cap (e.g., $5, $10, $50)
- Key enters the pool
- Sponsor receives monthly impact report (anonymized stats only)
Supported providers: Any OpenAI-compatible API — Gemini, DeepSeek, OpenAI, Groq, OpenRouter, etc.
5.2 Key Pool Management
- Round-robin or remaining-quota-based distribution
- Auto-failover when a key hits its cap
- Sponsor can pause/resume/remove key anytime
- Alert sponsor when approaching budget limit
5.3 Security
- Sponsor keys stored encrypted (AES-256) on proxy server
- Keys NEVER exposed to user devices
- Conversation content NEVER sent to sponsors
- Sponsors cannot identify individual users
- Proxy strips all PII before cloud API call
5.4 Revenue Paths (Platform Sustainability)
| Path | Description |
|---|---|
| Sponsor donations | Individual developers donate API keys |
| Corporate CSR | Bulk sponsorship (e.g., 1,000 users/month = ~$100) |
| B2G | Government procurement (지자체, 아동권리보장원) |
| Premium tier | Power users pay for own cloud access |
| Grants | Social impact grants (아름다운재단, etc.) |
Rule: Platform NEVER handles money for users. Only API key intermediation.
6. Crisis Management
6.1 Detection
- Keyword monitoring: suicide, self-harm, abuse, fraud indicators
- Behavioral signals: sudden inactivity after distress, repeated crisis-adjacent topics
- Sentiment shift detection
6.2 Response Protocol
[Crisis Detected]
→ AI IMMEDIATELY stops normal conversation
→ Displays: "지금 많이 힘들구나. 혼자 감당하지 않아도 돼."
→ Shows ONE-TAP emergency buttons:
┌─────────────────────────┐
│ 🆘 자살예방상담 1393 │
│ 📞 정신건강위기 1577-0199│
│ 🚨 긴급신고 112 │
│ 💬 계속 대화하기 │
└─────────────────────────┘
→ AI does NOT attempt to counsel
→ AI does NOT diagnose
→ AI stays present: "나는 여기 있어. 전화하기 어려우면 같이 있을게."
6.3 Night Mode (22:00–06:00)
- Increased crisis detection sensitivity
- Emergency buttons more prominent
- Tone shifts to extra-gentle
7. Data & Privacy
7.1 Local-First Principle
| Data | Location | Encryption |
|---|---|---|
| User profile | Device only | AES-256 |
| Chat history | Device only | AES-256 |
| Calendar data | Device only | AES-256 |
| Financial data | Device only | AES-256 |
| Sponsor API keys | Proxy server | AES-256 |
| Usage statistics | Proxy server | Anonymized |
7.2 What the Proxy Server Knows
- Timestamp of cloud API request
- Which sponsor key was used
- Token count consumed
- NOTHING ELSE — no user ID, no conversation content, no profile data
7.3 Device Migration
- Encrypted backup to Google Drive / iCloud
- QR code-based device-to-device transfer
- Backup file is unusable without device PIN/biometric
7.4 Data Deletion
- User can delete all data anytime (Settings → "모든 데이터 삭제")
- No server-side data to delete (nothing is stored)
7.5 Minors Protection
- Users under 18: adjusted tone, stricter crisis response
- Facility staff connection option (for in-care youth)
- Legal guardian consent flow (noting that some users have no legal guardian — design for this edge case)
8. LLM Stack
8.1 On-Device (Free Tier)
| Model | Size | RAM Required | Use Case |
|---|---|---|---|
| Gemma 4 E4B | ~9.4B total / 4B effective | 8GB+ | Primary on-device model |
| Gemma 4 E2B | ~5.1B total / 2.3B effective | 6GB+ | Low-end device fallback |
Capabilities: Text + Image + Audio native, 128K context, reasoning mode
8.2 Cloud (Sponsor Tier)
| Priority | Provider | Model | Cost/1M tokens | Notes |
|---|---|---|---|---|
| 1 | Gemini Flash | Free tier generous | Best Korean, multimodal | |
| 2 | DeepSeek | V3.2 | ~$0.14 input | Cheapest high-quality |
| 3 | Alibaba | Qwen 3.5 | Low | Best CJK vocabulary |
| 4 | OpenAI | GPT-5 Nano | Low | Via OpenRouter free |
| 5 | Any | OpenAI-compatible | Varies | Sponsor's choice |
8.3 Model Routing Strategy
User query arrives
├─ Simple (greeting, daily chat, emotional support)
│ → On-device Gemma 4 E4B (FREE)
│
├─ Medium (life advice, skill teaching)
│ → On-device Gemma 4 E4B (FREE)
│
└─ Complex (welfare eligibility, contract review, multi-step admin)
└─ Has sponsor access?
├─ Yes → Proxy → Sponsor API key → Cloud LLM + RAG
└─ No → On-device best-effort + "자세한 건 여기서 확인해" + link
8.4 Fine-Tuning Plan
- Domain: Korean administrative terminology, welfare system vocabulary
- Tone: Role-based persona tuning (parent/senior/friend modes)
- Safety: Crisis response patterns, fraud detection patterns
- Base: Gemma 4 E4B (Apache 2.0 allows commercial fine-tuning)
- Method: LoRA/QLoRA on 22B Labs PC (RTX 4080 Super 16GB)
9. Technical Stack
9.1 Client
- Framework: React Native (cross-platform iOS/Android)
- On-device inference: ExecuTorch (Meta) for production / llama.cpp for prototype
- Local DB: SQLite + SQLCipher (encrypted)
- Push notifications: Firebase FCM (proactive care triggers)
- Audio: Gemma 4 native audio pipeline (no separate STT/TTS)
9.2 Proxy Server
- Runtime: Fastify (Node.js) — lightweight, stateless
- API key vault: HashiCorp Vault or encrypted SQLite
- RAG: LangChain + government data sources
- Monitoring: Usage counters only (no content logging)
- Hosting: Single VPS initially (22B Labs PC for dev)
9.3 KakaoTalk Channel
- Kakao i Open Builder integration
- Routes to proxy server for cloud inference
- Fallback to template responses when all sponsor keys exhausted
10. Legal Considerations
| Law | Relevance | Action Required |
|---|---|---|
| 개인정보보호법 | User profile, chat data | Local-only storage; minimal server data; privacy policy |
| AI기본법 (pending) | AI service regulation | Monitor legislative progress; prepare compliance |
| 아동복지법 | Minors as users | Special consent flow; facility staff integration |
| 전자금융거래법 | Sponsorship model | Platform does NOT handle money — API keys only |
| 고독사예방법 | Elderly user safety | Crisis detection; emergency connection protocol |
11. Graduation Design
Gyeot is designed for users to outgrow it.
Signals
- Decreasing usage frequency
- User handling tasks independently that they previously asked about
- User forming real social connections (mentioned in conversation)
Response
- Acknowledge growth: "많이 컸다, 이제 혼자서도 잘 하네"
- Offer mentor transition: "네 경험이 누군가한테 도움이 될 수 있어"
- Reduce proactive check-in frequency (never abruptly stop)
- Always available if user returns
12. Roadmap
Phase 0: Foundation (Month 1–2)
- [ ] Repository setup (sinmb79/gyeot)
- [ ] On-device Gemma 4 E4B prototype (Android)
- [ ] Basic conversational onboarding
- [ ] Parent role persona implementation
- [ ] Local encrypted storage
Phase 1: Core Experience (Month 3–4)
- [ ] Proactive care system (meal check-in, weather, inactivity)
- [ ] Welfare benefit scanner (RAG with 복지로/자립정보ON)
- [ ] Sponsor API key proxy server
- [ ] Crisis detection & emergency connection
- [ ] KakaoTalk channel (basic)
Phase 2: Life Integration (Month 5–6)
- [ ] Calendar integration (Google/Apple)
- [ ] Simple financial tracker
- [ ] Multiple role personas (senior, friend, neighbor)
- [ ] Voice conversation (Gemma 4 native audio)
- [ ] Device migration (encrypted backup)
Phase 3: Community (Month 7–8)
- [ ] Sponsor dashboard & impact reports
- [ ] Growth tracking & graduation design
- [ ] Human mentor bridge (바람개비서포터즈 etc.)
- [ ] Multi-language support (다문화가정)
- [ ] iOS release
Phase 4: Scale (Month 9+)
- [ ] B2G pilot with local government
- [ ] Corporate CSR partnerships
- [ ] Fine-tuned domain model release (open-source)
- [ ] Community contributions & forks
- [ ] International adaptation framework
13. Metrics
Impact Metrics (Primary)
- Users who successfully completed an administrative task for the first time
- Welfare benefits discovered and applied for
- Crisis situations safely escalated to real services
- Users who "graduated" to independent life or mentor role
Operational Metrics (Secondary)
- Daily active users
- On-device vs cloud query ratio (lower cloud = better)
- Sponsor key pool health (keys available / keys exhausted)
- Average response latency
Anti-Metrics (What We Refuse to Optimize)
- Session duration (we WANT users to spend less time, not more)
- Engagement / retention (dependency is failure)
- Data collection volume (less is better)
Appendix A: Competitive Landscape
| Service | Target | Approach | Gap Gyeot Fills |
|---|---|---|---|
| Replika | General lonely users | Companion chatbot | No role structure, no practical life support |
| ElliQ | Elderly | Hardware robot | Expensive, English-only, no administrative help |
| 효돌 | Korean elderly | AI robot | Hardware-dependent, no youth coverage |
| 다솜 | Korean elderly | Smart speaker | Elderly only, no role-based relationship |
| 서울톡 | Seoul citizens | Admin chatbot | Information-only, no emotional relationship |
| ChatGPT/Claude | General | Generic AI | No proactive care, no role personas, no welfare integration |
Gyeot's unique position: Role-based relationship × Practical life support × Proactive care × Local-first privacy × Sponsor-funded access × Designed for graduation
Appendix B: Key References
Research
- Stanford IRB Study: "Loneliness and suicide mitigation for students using GPT3-enabled chatbots" (npj Mental Health Research, 2024)
- STAT News: "Prosthetic relationships" framework for AI companions (2025)
- 한국사회보장정보원: "가족돌봄청년(Young Carer) 기초 연구" (2024)
- 보건복지부: "2023 자립지원 실태조사" (2024)
- 보건복지부: "2024년 고독사 발생 실태조사" (2025)
- 성평등가족부: "2024년 한부모가족 실태조사" (2025)
- 국무조정실: "2024년 청년의 삶 실태조사" (2025)
Policy
- 아동복지법 제38조 (자립지원)
- 고독사 예방 및 관리에 관한 법률
- 청년기본법
- 한부모가족지원법
Technical
- Google Gemma 4 (Apache 2.0, April 2026)
- Off Grid: Open-source on-device LLM app (MIT)
- ExecuTorch: Meta's mobile inference framework
- awesome-mobile-llm: github.com/stevelaskaridis/awesome-mobile-llm
곁 — 너의 곁에서, 너와 함께.
22B Labs · The 4th Path
P4 := ⟨H⊕A⟩ ↦ Ω