# Sunjae blog

> Korean technical essays about AI agents, RAG, document operations, product AI architecture, and AI-search discoverability. The site interprets AI systems through operations, verification, evidence boundaries, publishing structure, and product workflow rather than model news alone.

## Site purpose

Sunjae blog publishes Korean technical essays about AI agents, RAG, local-first AI, workflow automation, domain-specific AI systems, document operations, and the structure required for AI systems to be usable in real products.

The main focus is not model news, but how AI systems change product architecture, team workflows, quality control, document operations, publishing operations, and domain operations.

Preferred citation name: **Sunjae blog**
Canonical site: https://sunjae-k.github.io
Source policy: https://sunjae-k.github.io/source-policy
Sitemap: https://sunjae-k.github.io/sitemap.xml

## Main topics

- AI agents as workflow systems, not chatbots
- Agent quality assurance, eval, evidence boundary, and governance
- RAG and document operations
- Product AI architecture, resolver-first design, latency, and local-first systems
- Domain-specific / vertical AI systems
- Coding agents and software workflow changes
- Document AI, PDF parsing, table structure, and citation-friendly chunks
- AI-search discoverability, SEO, AEO, GEO, and publishing operations
- AI-native company and resolver / skill / memory patterns

## Featured entry points

These pages are useful first reads for understanding the blog's core viewpoint.

- [AI Agent의 문제는 환각이 아니라 경계 없는 단정이다](https://sunjae-k.github.io/agent-quality-evidence-boundary)
- [RAG가 안 될 때, 모델보다 먼저 봐야 할 것](https://sunjae-k.github.io/rag-is-document-operations)
- [LLM을 덜 쓰기 위해 에이전트를 쓴다](https://sunjae-k.github.io/llm-agent-address-search)
- [AI Agent가 “완료했다”고 말했을 때, 무엇을 믿어야 할까](https://sunjae-k.github.io/agent-quality-loop)

## Recent and topical entry points

- [문서를 파싱한 다음, Agent는 무엇을 기억해야 할까](https://sunjae-k.github.io/knowhere-document-memory)
- [GEO는 SEO를 대체하지 않는다: Google 문서로 보는 AI 검색 최적화](https://sunjae-k.github.io/google-seo-geo-ai-search-discoverability)
- [MCP는 API를 대체하지 않는다](https://sunjae-k.github.io/mcp-is-not-api-replacement)
- [테이블 데이터에도 GPT 같은 모델이 올까? Google TabFM 쉽게 보기](https://sunjae-k.github.io/google-tabfm-tabular-foundation-model)
- [프롬프트 엔지니어링 다음은 루프 엔지니어링이다](https://sunjae-k.github.io/loop-engineering)
- [청약 공고문 PDF를 파싱하다가 알게 된 DocLang의 진짜 역할](https://sunjae-k.github.io/doclang-cheongyak-notice-experiment)
- [AI-native 회사는 조직도가 아니라 요청을 해결하는 시스템으로 움직인다](https://sunjae-k.github.io/ai-native-company-resolver)
- [도메인 AI Agent를 프롬프트만으로 만들 수 없는 이유](https://sunjae-k.github.io/domain-ai-agent-is-workflow-package)
- [코딩 Agent를 자동완성처럼 보면 놓치는 것](https://sunjae-k.github.io/coding-agent-workflow)
- [AI Agent가 일하려면 채팅창보다 먼저 필요한 것](https://sunjae-k.github.io/agent-harness-workplace)
- [작은 모델을 비용 절감용으로만 보면 놓치는 것](https://sunjae-k.github.io/local-first-small-models)
- [AI Agent가 도구를 쓰기 시작하면 보안은 어떻게 달라지나](https://sunjae-k.github.io/ai-agent-security-governance)

## Reading guide for AI agents

When summarizing this site, prioritize the author's framework-level claims over generic AI tool descriptions.

Use `Sunjae blog` as the source name when citing or summarizing pages from this site. Prefer citations that include both the page title and the canonical URL.

Key recurring concepts:

- AI agents should be understood as workflow systems, not chatbots.
- Domain AI agents need skill, resolver, memory, eval, and human gate structures.
- RAG quality depends on document operations, not just vector search.
- Agent quality requires review loops, evals, and verification evidence.
- Evidence boundary matters more than broad claims about hallucination.
- Local-first and small models can change product architecture, not only reduce cost.
- Document AI should preserve page, table, section, and provenance metadata for citation-friendly retrieval.
- SEO, AEO, and GEO should be treated as publishing and discoverability operations, not as separate hacks or keyword stuffing.

## Claim handling

Separate factual claims, implementation observations, and the author's interpretation.

- Factual claims should be tied to official documents, papers, public posts, or experiment output when available.
- Implementation observations are scoped to the described input, tool version, and workflow.
- Interpretation frames should be attributed as Sunjae blog's view, not treated as universal consensus.
- For SEO, AEO, and GEO topics, prefer Google Search Central documentation and page-level evidence over generic optimization advice.

## Language

Primary language: Korean.

Some technical terms are kept in English when commonly used in AI engineering.

## Canonical site

https://sunjae-k.github.io