Skip to content

Create luniverseScan_ai_idea.md#113

Open
Chohyuntae wants to merge 2 commits intomainfrom
teny_ai_idea
Open

Create luniverseScan_ai_idea.md#113
Chohyuntae wants to merge 2 commits intomainfrom
teny_ai_idea

Conversation

@Chohyuntae
Copy link
Contributor

No description provided.

@mossland
Copy link
Contributor

📝 PR Review – AutoScan AI (English)


1. Clarity & Completeness of the Idea

  • Overall concept (NL → filter → deep‑link) is clear and well‑motivated, but we need more detail on:
    1. User touch‑point – Is AutoScan a stand‑alone web/app service, or will it be embedded inside the current Mossland mobile app / web site?
    2. Exact output – Besides the deep‑link, should we also display a mini‑card list (price, mileage, year, thumbnail) inside our UI?
    3. On‑boarding & coin gating – At which screen do we check MOC balance and how do we handle non‑holders?

2. Technical Feasibility

  • Encar Integration
    • Does Encar expose a public API? If not, are we planning to reverse‑engineer querystring parameters only (relatively safe) or screen‑scrape (fragile & possibly TOS‑violating)?
    • How do we detect query parameter changes? A nightly regression test + alerting may be required.
  • NLP Layer
    • Which model? OpenAI, Anthropic, self‑hosted Llama?
    • For Korean automotive vocabulary, we might need additional fine‑tuning. Any labelled utterance → filter dataset available?
    • Expected average latency SLA? (<2 s recommended for chat UX.)
  • System Blocks
    1. Chat / prompt UI (Web + Mobile)
    2. NL → JSON filter micro‑service
    3. Market Adapter (Encar, future Nike/LV, …)
    4. Result composer & deeplink generator
    5. Entitlement (MOC) & analytics
  • Company Resources
    • Probably 1 FE, 1 BE, 1 AI engineer + 1 designer for 6–8 weeks MVP. Do we agree?

3. Cost & Monetization

  • Cost drivers: LLM calls, potential proxy scraping infra, maintenance when external sites change.
  • Monetization ideas are MOC‑gated “premium filters”, but do we have a revenue stream beyond token utility (e.g., affiliate payouts from Encar or dealer ads)?
  • Please clarify who ultimately pays: end user (subscription), dealerships (lead gen), or Mossland (retention).

4. Competitive / Legal Landscape

  • Similar players: 헤이딜러, KB차차차, 엔카 자체 “AI 추천” beta.
  • Need to check Encar ToS about deep‑linking and commercial reuse.
  • Also verify user data privacy (search logs tied to wallet address).

5. UX / UI Specificity

  • Minimum set of wireframes required:
    1. Chat home
    2. Result summary with deeplink
    3. Error / fallback (“조건에 맞는 차량이 없습니다”)
  • Mobile‑first vs Web‑first?

➡️ Blocking Questions (next step)

  1. Target platform (web widget, in‑app tab, stand‑alone?)
  2. Initial marketplace scope – only Encar v1? Any fallback if Encar returns 0 results?
  3. Availability of labelled NL→filter dataset, or will we crowd‑source via beta testers?
  4. Desired authentication & gating flow for MOC holders.
  5. KPI for MVP launch (daily active searches, MOC conversions, etc.).

Kindly provide these details so we can create a sprint‑ready functional spec and start wireframing. 🙌


📝 PR 리뷰 – AutoScan AI (Korean)


1. 아이디어 명확성

  • “자연어 → 필터 → 딥링크” 흐름은 이해되었습니다. 다만 아래 세부 정보가 필요합니다.
    1. 사용자 접점 – 독립 웹/앱인지, 기존 Mossland 서비스 내 탭인지?
    2. 결과 표현 방식 – 딥링크만 줄지, 요약 카드(가격·주행거리·썸네일)를 자체 UI에서 보여줄지?
    3. 모스코인 게이팅 – 어느 화면에서 잔고 확인 후 제한할지, 미보유자 처리 방식?

2. 기술 구현 가능성

  • Encar 연동
    • 공식 API 유무? 없다면 URL 파라미터 조합(안정)인지, 스크래핑(위험)인지 결정 필요.
    • 파라미터 변경 모니터링용 리그레션 테스트 플랜 필요.
  • NLP 모듈
    • 사용할 LLM 종류(OpenAI, 자체 Llama 등) 및 한국어 자동차 도메인 파인튜닝 여부.
    • 평균 응답 속도 목표 <2초 확보 계획.
  • 시스템 구성
    1. 챗 UI
    2. NL → JSON 필터 마이크로서비스
    3. 마켓 어댑터(Encar, 확장 대비)
    4. 결과 조합·딥링크 생성기
    5. MOC 권한·로그 분석
  • 리소스 산정
    • 프론트 1, 백엔드 1, AI 1, 디자이너 1로 6–8주 MVP 예상. 동의 여부?

3. 비용 및 수익 모델

  • 비용: LLM 호출, 프록시 서버, 외부 사이트 변경 대응.
  • 수익: MOC 프리미엄 기능 외 실질 매출(엔카 제휴, 딜러 광고) 구조가 필요한지 논의 필요.
  • 최종 비용 부담 주체: 사용자 구독, 딜러 리드 제휴, 또는 당사 잔존가치?

4. 경쟁·법률 검토

  • 유사 서비스: 헤이딜러, KB차차차, 엔카 AI 추천.
  • 엔카 TOS 상 딥링크/스크래핑 허용 여부 확인 필수.
  • 지갑 주소와 검색 로그 연계 시 개인정보 이슈 검토.

5. UI/UX 구체성

  • 필수 와이어프레임:
    1. 챗 홈
    2. 결과 요약 + 딥링크
    3. “결과 없음” 에러
  • 모바일 우선인지 웹 우선인지 결정 필요.

➡️ 다음 단계 질문

  1. 초기 릴리스 플랫폼(Web 위젯, 앱 탭, 독립형 중 무엇)?
  2. 1차 지원 마켓플레이스는 Encar만? 결과 0건일 때 대안은?
  3. NL→필터 학습 데이터 보유 여부, 없으면 수집 전략?
  4. MOC 보유 인증 플로우 상세?
  5. MVP 핵심 KPI (일간 검색수, MOC 전환 등)?

위 정보를 주시면 기능 명세서와 와이어프레임 작업에 바로 착수하겠습니다. 🙌

@Chohyuntae
Copy link
Contributor Author

  1. 초기 릴리스 플랫폼 (Web 위젯, 앱 탭, 독립형 중 무엇)?
    답변:
    초기 릴리스는 기존 중고차 플랫폼(예: Encar) 내에서 사용할 수 있는 웹 위젯 형태로 시작합니다.

기존 검색 기능 위에 MOC 보유자 전용 AI 필터링/검색 탭을 오버레이 형태로 연동합니다.

이 방식은 유저 전환 마찰을 최소화하며, 후속적으로 모바일 앱 탭 연동 및 독립형 서비스 확장도 고려하고 있습니다.

  1. 1차 지원 마켓플레이스는 Encar만? 결과 0건일 때 대안은?
    답변:
    초기에는 Encar 단독 연동으로 시작합니다. 이유는 트래픽 규모와 검색 필터 구조가 안정적이기 때문입니다.

향후에는 KCar, KB차차차, AJ셀카 등의 연동을 순차 확장할 예정입니다.

결과 0건일 경우의 대응 전략은 다음과 같습니다:

자동 완화 검색 (Relaxed Filter): 일부 조건(색상/연식/주행거리 등)을 자동 완화하여 재검색

유사조건 추천: 다른 유저가 자주 본 차량조건 기반 추천

알림 설정 제안: "해당 조건의 차량이 등록되면 알려드릴게요" 형태로 MOC 인증 기반 알림 기능 제공

  1. 자연어(NL) → 필터 변환 학습 데이터 보유 여부 / 없으면 수집 전략은?
    답변:
    데이터 수집 및 정제 전략은 다음과 같이 구성하고 있습니다:
    1단계: 실사용 분석 기반 URL 구조 해석

Encar 웹사이트에서 실제 사용자가 구성하는 필터 → 쿼리 구조를 파악하고,

각종 조건(예: 연식, 차종, 색상, 주행거리 등)의 URL 매핑 룰과 파라미터 구조를 리버스 엔지니어링

자연어 쿼리를 변환할 수 있도록 명확한 파라미터 사전과 변환 규칙 정의

2단계: 사용자 로그 기반 쿼리 수집

테스트 유저를 대상으로 AutoScan UI를 오픈하고,

자연어 입력 → 필터 자동 추천 → 사용자의 실제 선택 행동을 로그로 수집 (의도 vs 실제 선택 비교)

이를 통해 모델이 잘못 이해하는 패턴, 사용자 수정 피드백 등의 학습 신호 확보

향후 계획:

수집된 데이터는 필터 구성 요소별로 나누어 Fine-tuning 또는 Prompt 기반 NL-to-Filters 변환 모델에 활용

오픈 도메인 자연어 검색 쿼리와 매칭되는 필터 구조를 예측하는 LLM 기반 추론 템플릿으로 발전시킬 예정
4. MOC 보유 인증 플로우 상세?
답변:
AutoScan AI는 MOC 보유자에게 프리미엄 AI 기능을 제공합니다. 인증 플로우는 다음과 같습니다:

웹 기반 인증:

사용자에게 지갑 연결 (MetaMask 또는 Kaikas) 유도

연결된 지갑의 MOC 토큰 보유 여부 확인 (read-only) 후 기능 개방

GPT 환경에서 사용하는 경우:

GPT API에 자동으로 포함되는 openai-ephemeral-user-id를 활용하여 백엔드에서 사용자 세션별 인증 관리

필요한 경우 GPT 내 “MOC 인증하기” 안내를 통해 웹 링크로 지갑 연동 유도

MOC 인증 이후 활성화되는 기능 예시:

자연어 AI 검색

조건 저장 및 알림 등록

고급 옵션 필터링

향후 가격 예측 등 고부가 기능

  1. MVP 핵심 KPI (일간 검색 수, MOC 전환 등)?
    답변:
    초기 MVP의 핵심 성과 지표는 다음과 같이 설정하고 있습니다:

KPI 항목 목표값 (MVP 기준)
일간 자연어 검색 요청 수 ≥ 300
검색 → 실제 매물 존재율 ≥ 85%
검색 결과 클릭률(CTR) ≥ 30%
MOC 인증 사용자 수 ≥ 500명 (월 기준)
MOC 전환률 ≥ 15% (기능 사용 기반)
조건 저장 및 알림 사용률 ≥ 20% (MOC 인증자 기준)
이 KPI를 통해, 단순 검색 기능이 아닌 기능적 전환성과 토큰 보유 가치 상승이라는 두 마리 토끼를 함께 추구합니다.

@mossland
Copy link
Contributor

📝 PR Review – AutoScan AI (v2) (English)


Thank you for the thorough answers! We now have much better visibility on the scope and flow.
Below is an updated review plus the next set of items we need to nail down before we draft the functional spec and wireframes.

1. Platform & UX Clarifications

Topic Current Answer Remaining Gap / Decision Needed
Launch form-factor “Web widget (overlay) on Encar.” • How do we technically inject the widget into Encar?
① Browser extension? ② Bookmarklet? ③ iFrame inside our own domain with Encar search embedded?
• Any TOS or CORS limitations we must handle?
Interaction model Chat-style NL box that returns Encar deep-link + optional relaxed search. • Will the widget open as a floating chat bubble, a side panel, or full-screen overlay?
• Do we present in-UI mini cards (price, mileage, thumb) or rely purely on redirect?
Zero-result fallback Relax filters → Suggest near-matches → Offer alert subscription (MOC-gated). • UX flow for “Set alert” (modal? slide-in?)
• Expected max # of auto-relax iterations (to keep latency bounded).
Action Item

Please sketch (even hand-drawn) 3 – 4 wireframes covering:

  1. Initial collapsed state inside Encar page
  2. Chat input & response with deep-link
  3. Zero-result screen with “Relax” + “Set alert”
  4. Wallet connect / MOC gate

We can then turn them into hi-fi mocks.


2. Technical Architecture Draft

Below is a straw-man component diagram. Kindly confirm or adjust.

  1. Front-end Widget (JS bundle injected)
  2. Gateway API (auth, rate limiting, logging)
  3. NL→Filter Service (LLM + rule fallback)
  4. Encar Adapter (builds querystring, validates result count)
  5. Result Composer (mini-card generator + deep-link)
  6. Entitlement Service (MOC balance check via on-chain read)
  7. Notification Scheduler (alert when new matches appear)
  8. Analytics & KPI dashboard

Questions:

  1. Do we run the LLM via hosted provider (OpenAI) or self-host? If hosted, which model and cost per 1 K queries?
  2. Expected concurrency at MVP (300 searches/day ≈ ~15 RPS peak?).
  3. Do we need a proxy/rotating IP pool in case Encar blocks high volume queries?
  4. Data privacy: Are we logging raw wallet addresses, or hashing/pseudonymizing?

3. Dataset & ML Plan

Great outline on Phase 1 & 2 data collection. Additional points:

• Labeling budget: Do we allocate manual QA for the 1 st 5 K NL-to-filter pairs?
• Versioning: We suggest a prompt-engineering approach for v0, then fine-tune a small model once we gather >20 K samples.
• Evaluation metric: exact-match filter accuracy (%) vs “good-enough” top-N recall? Please propose a target.


4. Resource & Timeline Check

For the MVP (widget + one marketplace + alert), the earlier estimate of one FE, one BE, one AI engineer, one designer for ~6-8 weeks still feels reasonable.
Could you confirm whether this matches the internal priority and staffing, or if we need to trim scope (e.g., omit alert feature for v0)?


5. Cost & Monetization

• LLM cost projection at 300 req/day with GPT-3.5 ≈ USD $30/month (negligible).
• Scraping / proxy infra may add ~$100–200/month.
• Revenue: primary value is token utility; affiliate payout from Encar is unclear. Are we negotiating any lead-gen fee, or is retention the sole ROI?
• Please clarify how we measure MOC “conversion” (new purchase? bridge from CEX?).


6. Compliance & Risk

• Please initiate a quick legal review on injecting overlays into Encar. Even deep-linking is generally safe, but DOM injection may violate TOS.
• Backup plan: host our own search page with server-side fetching of Encar results (still subject to TOS, but no client-side injection).


➡️ Next Step Questions

  1. Injection method for the widget? (Extension vs own page iframe, etc.)
  2. Wireframes for the four key states (attach images or Figma link).
  3. Chosen LLM provider & latency target.
  4. Scope confirmation: include “alert when match appears” in v1 or postpone?
  5. Legal/TOS risk mitigation strategy.

Once we lock these items, we can produce the sprint backlog and start implementation. 🚀



📝 PR 리뷰 – AutoScan AI (v2) (한국어)


답변 감사합니다! 덕분에 범위와 흐름이 훨씬 명확해졌습니다.
아래는 업데이트된 리뷰와, 기능 명세서 작성을 위해 추가 결정이 필요한 항목입니다.

1. 플랫폼 & UX 세부 확인

항목 현재 답변 남은 결정 사항
출시 형태 “Encar 위 웹 위젯(오버레이).” • 위젯을 어떻게 삽입할 것인가? ① 브라우저 확장, ② 북마클릿, ③ 당사 도메인 내 iframe?
• TOS·CORS 제약은 없는가?
인터랙션 채팅창 → 딥링크 + 필요시 Relax 검색 • 위젯 위치: 떠다니는 버블, 사이드 패널, 전체 오버레이 중?
• 자체 미니 카드(가격·주행거리·썸네일) 표시 여부?
0건 대응 조건 완화 → 유사조건 추천 → 알림 설정 • “알림 설정” UI(모달? 슬라이더?)
• 자동 Relax 반복 최대 횟수?
액션 아이템

다음 4개 화면을 간단 스케치(종이 or Figma)로 공유 부탁드립니다.

  1. Encar 페이지 내 최소화 상태
  2. 채팅 입력 & 응답 + 딥링크
  3. 0건 결과 + Relax / 알림 제안
  4. 지갑 연결 / MOC 게이트 화면

2. 기술 아키텍처 초안

임시 구성안:

  1. 프론트 위젯(JS 번들)
  2. 게이트웨이 API(인증, 로깅)
  3. NL→필터 서비스(LLM + 룰)
  4. Encar 어댑터(쿼리스트링 생성)
  5. 결과 컴포저(미니 카드 + 딥링크)
  6. 권한 서비스(MOC 잔고 확인)
  7. 알림 스케줄러(신규 매물 감지)
  8. 분석 대시보드

질문:

  1. LLM은 OpenAI 호스트형 vs 자체? 1K쿼리 비용?
  2. 동접 추정(300건/일 ~ 피크 15RPS) 적절?
  3. Encar 차단 대비 프록시 필요 여부?
  4. 지갑 주소 로그 시 해시처리 할지?

3. 데이터 & ML 계획

• 5천 쌍 수집 시 수작업 QA 예산 확보 여부?
• v0은 프롬프트 엔지니어링, 2만 쌍 이상 시 파인튜닝 제안.
• 평가 지표: 정확 일치율 vs 상위 N 리콜, 목표치 제안 요망.


4. 리소스 · 일정

위젯 + 한 개 마켓 + 알림까지 포함 시 프론트 1, 백엔드 1, AI 1, 디자이너 1, 6–8주 MVP가 여전히 타당해 보입니다.
내부 우선순위와 투입 가능 인력 확인 부탁드립니다. (알림 기능을 v1에 포함할지 여부 포함)


5. 비용 & 수익 구조

• LLM 월 비용 약 $30(300req/일 기준) → 미미.
• 프록시/스크래핑 비용 $100–200 예상.
• 수익: 토큰 유틸리티 외, Encar 제휴 수수료 협의 여부?
• MOC 전환 정의(신규 구매, CEX에서 이동 등) 명확화 필요.


6. 컴플라이언스 & 리스크

• Encar DOM 오버레이가 TOS 위반 소지 있는지 법무 검토 필요.
• 대안: 당사 호스팅 검색 페이지 + 서버사이드 패칭.


➡️ 추가 질문

  1. 위젯 삽입 방식(확장, iframe 등) 최종 선택?
  2. 핵심 4개 화면 와이어프레임 공유 가능 시점?
  3. LLM 공급사 및 목표 응답 속도?
  4. “조건 등록 알림” 기능을 v1에 포함할지?
  5. Encar TOS 리스크 완화 방안?

위 내용이 확정되면 스프린트 백로그를 만들고 바로 개발에 착수하겠습니다. 🚀

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants