벤더 설명:
"ORB is a new product coming out of the MinIO AI branch. MinIO
has a smaller team called MinIO AI... where we're building...
AI based products."
조직:
벤더 양해:
"I wish I could show you ORB because it's pretty impressive, but
[we] couldn't get the environment stabilizing [in] Kubernetes"
상황:
벤더 설명:
"Because today, when business has a question, they say, I would
like to know this. So someone needs to build a SQL statement or
[write] Python to go and fetch the data."
현재 프로세스:
Business Question
↓
Data Analyst (60분 소요)
↓
SQL/Python 작성
↓
Analysis Report
문제:
벤더 설명:
"we're trying to democratize access to data lake... AI, if AI,
can actually read out of a catalog... and either do the traditional
query or the traditional Python program."
ORB 프로세스:
Business Question (자연어)
↓
ORB (AI)
↓
Automated Analysis
↓
Instant Answer/Application
효과:
벤더 설명:
"object storage has done for the industry is standardize access
to the data to one common API, which is S3"
효과:
벤더 비유:
"very similar to what the JDBC connector did 25 years ago. JDBC
standardized access patterns to database... iceberg, standardized
access patterns for data across multiple platforms."
JDBC (25년 전):
Iceberg (현재):
벤더 강조:
"now the iceberg is emerging as the winner of the open format table.
That means that now we're standardizing access pattern... So now
this is when the data becomes structured. This is where AI thrives"
Iceberg → AI 활용:
벤더 설명:
"You have a question, we'll get you the answer that's more aligned
to... AI writing SQL or Python."
기능:
예시:
User: "지난 분기 매출이 가장 높은 제품 Top 5는?"
ORB:
[AI가 자동 생성한 SQL]
SELECT product_name, SUM(revenue) as total
FROM sales
WHERE quarter = 'Q4-2025'
GROUP BY product_name
ORDER BY total DESC
LIMIT 5;
[결과 표시]
1. Product A: $5.2M
2. Product B: $4.8M
...
벤더 설명:
"once you have figured out or understood the data, maybe you want
to interact with data. So that's the second interface. We have chat"
기능:
예시:
User: "지난 분기 매출을 보여줘"
ORB: [차트 표시]
User: "이 중에서 한국 매출만 보여줘"
ORB: [필터링된 차트]
User: "전년 대비 증감률은?"
ORB: [증감률 분석]
벤더 강조:
"whatever you're asking for, ORB will actually build an application
for you, put it in a container and deploy to Kubernetes. That's
the core feature of ORB."
프로세스:
1. User 요청 (자연어)
2. ORB 분석
3. 코드 생성 (Python + React)
4. 컨테이너 빌드
5. Kubernetes 배포
6. User 확인/피드백
7. 수정 반복
8. 최종 배포
벤더 예시:
"business can come and say, I would like to have a dashboard that
forecast sales on a quarter to quarter basis, but allows me to see
previous [quarters]."
ORB 작업:
1. Iceberg 카탈로그 분석
- sales 테이블 찾기
- 컬럼: date, product, revenue
- 관계: product → category
2. View/Query 생성
SELECT
QUARTER(date) as quarter,
SUM(revenue) as total_revenue
FROM sales
GROUP BY QUARTER(date)
3. Python Backend 코딩
- Flask/FastAPI
- Query 실행
- Forecast 알고리즘
4. React Frontend 빌드
- Line Chart (분기별 매출)
- Trend Line (예측)
- Filter (제품, 기간)
5. Docker 이미지 생성
FROM python:3.11
COPY backend/ /app
COPY frontend/build/ /app/static
6. Kubernetes 배포
kubectl apply -f dashboard.yaml
7. User에게 미리보기
"이게 맞나요? 색상을 바꿀까요?"
8. 피드백 반영
User: "파란색으로, 월별도 추가해줘"
ORB: [코드 수정 → 재배포]
9. 최종 승인
User: "완벽해요!"
ORB: Production 배포
벤더 강조:
"building a data application... takes probably 20 minutes."
타임라인:
벤더 설명:
"every three to six months, we see a 10 times improvement on
[foundational] models... six months ago, [it] used to take like,
40 minutes an hour, so now it's only 20 minutes."
벤더 설명:
"when you start ORB, we first connect to the catalog, list [all]
the tables, and start going one table... sampling... data multiple
times, so that we can start understanding what each column does
and what it means."
작업:
1. 카탈로그 연결: Iceberg REST Catalog
2. 테이블 목록: 모든 테이블 리스팅
3. 데이터 샘플링: 각 테이블 여러 번 샘플링
4. 컬럼 이해: 각 컬럼의 의미 파악
벤더 설명:
"after we do all this sampling... we call those table summaries"
예시:
Table: customers
Columns:
- customer_id (INT): Primary Key, 1~100000
- name (STRING): Person names
- email (STRING): Email addresses
- country (STRING): Country codes (KR, US, JP)
- created_date (DATE): Registration dates
Summary:
"Customer master table with 95,432 records.
Contains customer demographics.
Main markets: Korea (65%), US (20%), Japan (15%)."
벤더 언급:
"the [bootstrapping] takes a while because we're sampling so much
data... For us, it takes like an hour, and we only have 53 tables"
예상:
벤더 설명:
"we start doing what we call a correlation analysis, to try to
discover if there's data columns with similar names in different
[tables]."
분석:
테이블 관계 발견:
Table: customers
- customer_id (Primary Key)
Table: orders
- customer_id (Foreign Key?)
- order_id
Table: payments
- order_id (Foreign Key?)
- payment_id
ORB 추론:
"customers.customer_id → orders.customer_id
orders.order_id → payments.order_id"
벤더 언급:
"even if the table name is [encrypted], we will still know there's
a relationship. But if not, we establish the correlation."
능력:
벤더 설명:
"we have like... a clustering of tables, where we say, these tables
seem to be related, this table seems to be related."
결과:
Cluster 1: E-commerce
- customers
- orders
- payments
- products
Cluster 2: Logistics
- shipments
- warehouses
- inventory
Cluster 3: Marketing
- campaigns
- clicks
- conversions
벤더 설명:
"we present that to the ORB operator. That's kind of like the
bootstrapping process... You can review what ORB is assuming...
and then correct it"
프로세스:
1. ORB 추론 결과 표시
"customers.customer_id → orders.customer_id?"
2. Operator 확인
✅ Correct: "맞습니다"
❌ Incorrect: "아니요, customer_code입니다"
3. ORB 학습
메모리에 저장 → 향후 반영
4. 반복
모든 관계에 대해
벤더 설명:
"you... can say this group was actually one, or this table is
actually connected this way"
시나리오:
ORB: "Table A, B, C는 별개인 것 같습니다"
Operator: "아니요, 모두 같은 그룹입니다"
ORB: "알겠습니다. 학습했습니다."
전통적 방식:
Requirements → Design → Code → Test → Deploy
(수 주~수 개월)
벤더 비전:
"the entire software industry coming to a paradigm [where] software
[is not] written anymore. Software is created on demand."
새로운 방식:
Natural Language Request → ORB → Application (20분)
예시:
Old Way:
"Dashboard 필요" → PM → Dev → QA → Deploy (2주)
ORB Way:
"Dashboard 보여줘" → ORB → Dashboard (20분)
피드백 루프:
"If the user says, I would like to change colors, I would like
to add functionality, [in] just plain language, ORB actually
builds [it]"
예시:
User: "색상을 파란색으로 바꿔줘"
ORB: [코드 수정 → 재배포] "이렇게요?"
User: "필터 추가해줘 - 제품별, 지역별"
ORB: [기능 추가] "이제 필터가 있습니다"
User: "완벽해!"
ORB: "Production 배포할까요?"
벤더 설명:
"right now, we're using [foundational model] vendors, because the
GPUs that we have... it's not enough."
하드웨어:
외부 사용:
벤더 언급:
"we... fine tune[d] our model... when the product was called Unica.
We were running our own fine tune[d] models... for Prompt Object,
we also fine tune[d] the model."
이전 제품:
벤더 전략:
"we're building ORB in a way that the provider... It's not important
right now. So we can test different ones"
유연성:
벤더 우려:
"every region will have different requirements regarding the models
that they are allowed to use [in the] US, Anthropic, OpenAI, xAI
are fair play. But... Europe's already locking down some of the models."
예상:
벤더 강조:
"ORB can only understand the data if it's in a [structured] table.
Right now we [only support] Iceberg."
현재:
이유:
벤더 확인:
"ORB is still like three, three to five months away from being
released."
타임라인:
Stage 1: 데이터 접근 민주화 (현재)
벤더 설명:
"ORB is all about democratizing access to data. That's right now.
The stage one is democratizing access to data."
기능:
Stage 2: 데이터 수집 (향후)
벤더 계획:
"the next stage is to focus on ingestion, because... data needs
to come in real[time], as it's being produced, we want to start
[writing to] iceberg."
기능:
Stage 3: Real-time 애플리케이션 (미래)
벤더 비전:
"we'll start looking into more of a real time applications...
we definitely want ORB to be able to infer on data that's coming in."
기능:
고객사 설명:
"they have a lot of... data analysts. And they have a lot of data...
pressure sensor data... image... numbers... What they're doing
right now is they don't really have a pipeline. They rather just
collect all these... data, and then analyze it."
현황:
고객사 희망:
"I have this specific [recipe] for this semiconductor manufacturing,
and then... I want to know everything about it. So I don't want
to do separate analysis for image. I don't want to do separate
analysis for... pressure sensor data."
목표:
고객사 설명:
"They have two fault detection processes... one is real time to
make sure that nothing's going wrong. And two is... at the end
of the process, has it gone wrong?"
1. Real-time FDC:
2. End-of-Process:
고객사 설명:
"when you build a wafer in the fab, it takes eight months to
finally make [it] as a product. And for that eight [whole] month
data will keep being accumulated"
특성:
고객사 설명:
"They have two ways of analyzing things. One is analyzing, comparing
the data and the rule... and second one [is] only use the data...
see if it's a failure or not."
Rule-based:
IF pressure > 100 THEN alert("압력 초과")
IF temperature < 50 THEN alert("온도 부족")
Data-driven:
AI 모델이 데이터만으로 불량 판단
(학습된 패턴 기반)
고객사 요구:
"they will have to... do a distributed analysis, because they
don't want the results to come out [in a] few months... They
want [it in] real time or... within the eight month time frame."
요구사항:
벤더 답변:
"right now, we're [working on] data that [is already in] an
iceberg [table]. So that means the data already reached the
iceberg table somehow."
가능:
불가능:
벤더 계획:
"the next stage is to focus on ingestion... as data needs to
come in real[time]"
가능 (향후):
고객사 질문:
"structured data... images, image[s]"
벤더 답변:
"image[s], we haven't got [that]... we dropped that and we
started again."
현황:
벤더 인정:
"domain specific knowledge is something we don't have. That's
why ORB is still kind of like a general product."
현재:
고객사 설명:
"they already have like, 1000s of legacy applications, because...
if you just look at the numeric data... from the fab [you] can't
even decode it, like, ORB will not know what they're looking at."
상황:
벤더 솔직:
"I wish we could connect it to existing applications... But...
we're trying to have you teach it by actually explaining the process."
현재 방식:
이유:
"we don't want to say, dump [in] your source code and we figure
[it] out... because [it'll] make ORB feel more confused like
you have some code that decided to read this table... 100 times."
벤더 비유:
"it's [like] someone starts teaching like a child. Slowly, we
start by learning and mastering the process"
프로세스:
1. ORB 부트스트래핑: 테이블 이해
2. Operator 교정: 관계 수정
3. 점진적 학습: 반복
4. 내부 메모리: 축적
Internal Memory System:
"in the ORB internal memory system, it's trying to build like
an understanding of what each table, each relationship, is."
벤더 전망:
"I wouldn't be surprised that we'll be talking about it like in
a year or two, or even less."
예상:
벤더 언급:
"ORB may actually decide to train models... But right now...
we haven't even started [on] that. We only have talk."
현재:
벤더 설명:
"if you have the data, most of the problems that you can do on
top of data are known, and the most foundational model[s]...
are trained for that task."
예시:
User: "판매 예측 모델 만들어줘"
ORB (미래):
1. 데이터 분석 (sales 테이블)
2. 적합한 모델 선택 (Time Series Forecasting)
3. 학습 데이터 준비
4. 모델 학습 실행
5. 평가 및 배포
6. 예측 API 제공
어려움:
"it's hard... for me, it's hard to imagine that my product could
decide to train a very sophisticated model."
벤더 언급:
"one of the interesting features... is sandboxing the AI, because
sometimes the AI... tr[ies] some wild things."
이유:
벤더 설명:
"we don't let the AI run wild, but we put it in a isolated space
where [it] doesn't see other than the files it has"
격리:
데모 실패 이유:
"that's the part of our demo that failed... the sandboxing...
we're [still] trying to get it right"
상태:
ORB 안정화
문서화
이미지 지원 재개
GA 출시
Stage 2 계획
데이터 준비
요구사항 명확화
ORB 평가
Pilot 프로젝트
통합 계획
분석 시간:
애플리케이션 개발:
전문성:
병목 제거:
탐색적 분석:
문서 버전: 1.0
최종 수정일: 2026년 2월 4일
다음 리뷰: ORB GA 릴리스 후 (2026년 5~7월)
"we're trying to democratize access to data lake... AI... can
actually read out of a catalog... and either do the traditional
query or the traditional Python program."
"the entire software industry coming to a paradigm [where] software
[is not] written anymore. Software is created on demand."
"building a data application... takes probably 20 minutes."
"very similar to what the JDBC connector did 25 years ago...
iceberg, standardized access patterns for data"
"ORB... first connect to the catalog, list [all] the tables...
sampling... we call those table summaries... correlation analysis...
clustering of tables"
"domain specific knowledge is something we don't have. That's
why ORB is still kind of like a general product."