LLM WikiAccess-protected knowledge portal

WIKI

2026-W30 Worklog

AI Summary Purpose Tracks development work across repositories for this week. Key points 2026 07 28 User confirmed that the selected DuckDB take home package for Remember & Company Data Engineer was submitted. Current status take home submi

경로ai/worklog/2026/2026-W30.md
카테고리Worklog
태그#airflow #career-transition #cicd #crawler #human-output #infra #kubernetes #mysql #portfolio #study #w30 #worklog

# 2026-W30 Worklog

AI Summary

Purpose:

Key points:

Remember & Company Data Engineer was submitted. Current status: take-home submitted, awaiting the next result.

package. A fresh ZIP extraction completed ./up.sh with 2022/2023 SUCCESS, 32,687 master rows, zero source/master mismatches, a host-exported DuckDB, and 108 passed container tests. The gstack-style technical-lead score improved from 54/100 to 96/100 with no major blocker. The Korean work explanation received a separate independent technical-lead review and improved from 91/100 to 99/100. The 52-entry DB-free archive SHA-256 is 54b6c94210baa2ec347c0ea04ef35cf9200becfd59b74306e8056c100afe89e5.

human/study/content/ai-frontier/40-google-adk-2-5-remote-mcp-agent-to-mcp-cloud-run-sandbox.md, updated the ai-frontier series/curriculum control plane, and verified tests/build/relative links. The chapter frames ADK 2.5 as an execution- boundary release: exporting agents as MCP servers, consuming remote MCP server-side via ManagedAgent, and isolating code execution inside Cloud Run sandbox.

under human/study/content/database-frontier/36-materialize-v26-33-read-committed-control-plane.md, updated the database-frontier series/curriculum control plane, and verified tests/build/relative links. The chapter frames v26.33 as a control-plane contention release: PostgreSQL metadata DB READ COMMITTED consensus queries, session-scoped catalog snapshot caching, timestamp-oracle stall isolation, and replica-targeted EXPLAIN ANALYZE via the MCP developer endpoint.

source-backed ownership, concrete verification steps, and human Korean wording. Final local artifacts are a three-page resume and a 12-page portfolio with working full-card navigation; the deterministic document quality gate is 95.3/100 and all three independent reviews are at least 97.

human/study/content/ai-frontier/38-ai-sdk-7-workflowagent-tool-approvals-mcp-apps.md, updated the ai-frontier series/curriculum control plane, and verified tests/build/relative links. The chapter frames AI SDK 7 as an agent runtime surface shift (WorkflowAgent durability, runtime/tool context split, MCP Apps visibility separation, and explicit approval/timeout/sandbox contracts).

AI데이터부문 개발 프로젝트 담당 (Remember posting 327461, deadline 2026-07-29). Resume 3p PDF + portfolio 12p PDF finalized at ~/hw/클로드이력서/gs리테일-ai데이터부문/. This posting pivots to AI: the differentiating requirement is Codex/Claude Code development + Harness Engineering, so the resume front-loads AI-harness evidence instead of the usual DE ordering.

Elasticsearch index doc-count claim) from 클로드이력서/README.md.

evidence to the user-verified service boundary: AWS EC2 only, with MySQL self-managed on EC2. Removed RDS, S3, and EFS claims.

Relevant when:

application documents.

Do not read full document unless:

pipeline steps for the GS Retail application.

Linked documents:

Open Questions

user decision pending (default omitted, same pattern as the Toss Place dbt line). If accepted, regenerate the resume PDF.

Details

2026-07-26 — Remember Data Engineer take-home, Codex alternative

Agent: Codex. Standalone repo: /Users/khw/Library/CloudStorage/[email protected]/내 드라이브/projects/remember-data-engineer-assignment-codex.

This implementation intentionally explores a different policy and architecture from the Claude package:

Actual Airflow results from a clean database:

Verification and review:

Cross-Mac handoff saved for the next session:

2026-07-26 — Remember & Company Data Engineer take-home assignment

Agent: Claude Code. Working copy /Users/khw/hw/remember-de-assignment, delivered package career/클로드이력서/리멤버앤컴퍼니-data-engineer/과제/ (과제-원문.pdf, 김현욱_리멤버앤컴퍼니_DE과제.zip, 작업본/).

Build was done on a local disk rather than inside Google Drive because Docker bind mounts against the CloudStorage FileProvider path are slow and unreliable on macOS. Only the finished package is copied to Drive.

What the data actually contained. Both CSVs were profiled in full before any code was written, and the profile changed the design:

the assignment's "update with the 2023 values" literally destroys them.

numeric normalization, string comparison marks all 21,087 overlapping rows as changed.

formatting-only (시도 abbreviation and double spaces), 4 truncations, 4 refinements.

both years), 17 성립일자 moving backwards, 1 date beyond the source year, 30 U+3000 ideographic spaces in 사업장명.

Design decisions taken where the spec was silent, all collected in SyncPolicy and switchable by environment variable:

  1. Empty source value preserves the existing master value (COALESCE

semantics). Justified by the assignment text itself — the data is "일부를 추출하여 구성한 자료", so a blank means not extracted, not deleted. Running with the policy off raises master address nulls from 156 to 2,249.

  1. The 9,426 rows absent from 2023 are flagged is_active=false with

absent_since_year, never physically deleted. The spec defines no Delete rule, and SyncPolicy rejects a delete mode outright.

  1. Quality rules are graded ERROR (excluded from master) / WARN (applied and

recorded) / INFO (counted). Real runs produce zero ERRORs.

  1. One year-parameterized DAG (@yearly, catchup=true, max_active_runs=1)

covers both initial load and sync; initial load is just a sync against an empty master. Dropping a 2024 file into data/ needs no code change, and a missing year skips rather than fails.

  1. Change detection compares columns with IS DISTINCT FROM instead of a row

hash, so the rule is not implemented twice across Python and SQL.

Verified results. 2022 initial load 30,513 inserts; 2023 sync 2,174 inserts, 2,903 updates, 18,184 no-ops, 9,426 deactivations, 2,095 values preserved by the null policy; master 32,687 rows with 23,261 active, matching the 2023 source row count exactly. Every number reproduces the independent pre-implementation profile. Seven post-merge reconciliation checks pass on both runs, including a full-row comparison of policy-expected values against actual master values. Re-running a year is a no-op (0 inserts, 0 updates, 23,261 unchanged, no new history rows). 58 pytest tests pass. The DAG was run end to end under docker compose (Airflow 2.10.5 + Postgres) and produced results identical to the CLI path; the 2024 and 2025 runs skip as designed. A clean unzip into a fresh venv reproduces both the tests and the pipeline.

One defect was found and fixed during Docker verification: start_date anchored to Asia/Seoul while Airflow passes data_interval_start in UTC, which created a spurious 2021 run and misaligned each run label with the file year it loaded. Anchoring the DAG to UTC leaves exactly two real runs.

Limits recorded in the README rather than hidden: the 17 backwards 성립일자 are flagged but not adjudicated; the 4 truncated addresses are not caught because they are non-empty; duplicate-key handling is fixture-tested only since neither file has duplicates; address normalization stops at the 시도 level.

2026-07-26 — database-frontier study publication: OceanBase 4.6 distributed-SQL boundary release

Agent: Hermes cron (06:00 KST) in this repo.

Published one dynamic-fallback database/data-platform chapter:

Topic selection rationale:

Control-plane updates:

Verification run required by the study pipeline:

Expected commit message after staging/publishing flow:

2026-07-26 — database-frontier study publication: TiDB 8.5.7 hidden-waste control release

Agent: Hermes cron (05:00 KST) in this repo.

Published one dynamic-fallback database/data-platform chapter:

Topic selection rationale:

Control-plane updates:

Verification run required by the study pipeline:

Expected commit message after staging/publishing flow:

2026-07-25 — ai-frontier study publication: Agno 2.8 evaluation pipeline

Agent: Hermes cron (04:00 KST) in this repo.

Published one dynamic-fallback AI chapter:

Topic selection rationale:

Control-plane updates:

Verification run required by the study pipeline:

Expected commit message after staging/publishing flow:

Engineering review pass (2026-07-26, later same day). Ran the gstack plan-eng-review skill over the whole package, then a Codex outside-voice pass. Thirteen findings, all fixed and re-verified. Canonical working directory moved to Google Drive at the user's instruction (career/클로드이력서/리멤버앤컴퍼니-data-engineer/과제/작업본/), with the venv kept on local disk so it does not sync.

Outdated: the test counts, archive metadata, and unsplit merge.py decision in this point-in-time section were superseded by the 2026-07-27 final technical-lead review at the end of this worklog.

The load-bearing outcome reverses an earlier decision: null-preservation moved from default to option, and spec-literal became the default. The reasoning is that a take-home may be graded by diffing the master against the 2023 file, in which case 2,095 preserved values read as spec violations no matter how well the judgment is argued. Under the new default the master matches the 2023 source exactly (verification query returns 0 mismatches), while the README still leads with the 2,093-address loss analysis and recommends the flag for production. Both policies pass all reconciliation checks.

Defects found by review and fixed (each reproduced before fixing):

the expected value resolved to the master's own value, so the check could never fail for exactly the rows the headline decision governed. Added NO_SILENT_NULL_OVERWRITE, which proves the same property from the change log (recorded before the master is touched), and documented the limitation inline.

so a row with a broken 상시근로자수 dropped out of the merge, never advanced last_seen_year, and got deactivated as absent — a business physically present in the file marked as gone. With the 1% error gate that was up to 232 rows. Split into has_valid_key / is_valid; absence is now judged on the former.

0123456789 — manufacturing a key, colliding with a real one, and making the PK_FORMAT rule unreachable. Now left as-is and quarantined.

regressed last_seen_year, detected only by a post-hoc check that fired after the damage. Added assert_year_is_not_regressive() before the merge.

failure lost the previous snapshot while the DDL comment claimed "append-only". Wrapped in one transaction; comment corrected.

2022, headcount updated from 2023), which broke the advertised raw-trace query. Renamed to last_changed_source_year, documented as row-level only, and the trace query rewritten against master_change_log.

Removed, along with two comments that described uses which did not exist.

earlier ops_sync_run records. Now scoped to run_id.

FAILED alongside the previous success's row counts. Counters now reset too.

would fail on Linux hosts (container uid 50000 vs host ownership) — the recommended run path was broken for anyone not on Docker Desktop. Added user: "${AIRFLOW_UID:-50000}:0" plus .env.example.

project name from the directory, and a Korean-only name leaves no legal characters ("project name must not be empty"). The submission's own top-level folder is Korean, so this would have hit the evaluator. Pinned name: remember-de-assignment.

skipped locally (no Airflow) and did not exist in the container — they ran nowhere. Mounted, and pytest added to the image.

false, so the Docker path and the CLI path produced different results from the same data (address nulls 156 vs 2,249). Fixed, and locked with tests/test_config_consistency.py, which asserts the policy defaults agree across SyncPolicy, docker-compose.yaml, and .env.example.

Test coverage grew from 58 to 80. Local runs 74 passed / 1 skipped (DAG module needs Airflow); the container runs 69 passed / 11 skipped (repo-layout tests need the compose file). The two environments together execute every test at least once. Added regression tests for each reproduced defect, a negative test proving reconciliation actually fails on a corrupted master, and DAG contract tests covering the timezone regression.

Verified end to end after the fixes: CLI run, Docker Compose run from the Drive path (4 DAG runs — 2022/2023 succeed, 2024/2025 skip), container test suite, and a clean unzip into a Korean-named directory followed by both a venv run and a full Docker run. Final package: 김현욱_리멤버앤컴퍼니_DE과제.zip, 39 entries, 1.6 MB, UTF-8 filename flags set on every entry, no DuckDB file, no .env, no venv. Codex finding #2 ("ships a populated database") was checked and refuted for the artifact — the zip contains only warehouse/.gitkeep.

Not changed, by decision: merge.py still holds six pipeline stages in one module. Splitting it was judged a worse trade than the churn two days before the deadline, and that choice is stated in README §9 alongside the scale ceiling (stage() materializes a whole year in memory).

Cross-machine handoff (2026-07-26 EOD). Made the Claude package resumable from any Mac. The whole path is Korean, and macOS stores Korean filenames as NFC or NFD depending on how they were created, so a path copied from one machine can fail to resolve on another — the same problem Codex hit and solved with an ASCII lookup. Added the ASCII marker REMEMBER_CLAUDE_WORKDIR plus resume-here.sh, which finds itself, creates the virtualenv at ~/.venvs/remember-de-assignment (outside Drive, so it does not sync and does not carry non-portable interpreter paths), installs pinned requirements, then runs the tests, the pipeline, and the result query. --docker adds the full container E2E. Verified from scratch after deleting the venv: reproduces 30,513 / 2,174 / 4,000 / 17,087 / 9,426 and master 32,687 exactly. HANDOFF.md next to it carries the status, the expected-value table, the zip rebuild recipe, and the three environment traps.

Deleted the regenerable 27 MB DuckDB file from Drive; the working folder is now 7 MB. Registered the implementation in ai/repo-notes/index.md and ai/workspace/repos.md, wrote ai/repo-notes/remember-data-engineer-assignment-claude.md, and extended ai/workspace/machines.md with the marker-based lookup and the rule that virtualenvs and generated databases never live in Drive.

Corrected a now-false statement in the Codex repo note and in ai/workspace/active-context.md: both said the Claude alternative preserves previous values. After today's default reversal both implementations default to the literal update rule, so the remaining differences are DB engine (DuckDB vs SQLite), DAG count (one parameterized vs two explicit), and 2022-only row handling (is_active soft flag vs left untouched). Tomorrow is a comparison and submission decision, not implementation work.

One more submission defect fixed while packaging: reports/run-result.md embedded an absolute path containing the user's Google account address, because run_pipeline.py logged the DuckDB path verbatim. A pipeline that ships its own run log as a deliverable should not print absolute paths. Added _display_path() to log project-relative paths and regenerated the report; a grep for /Users/ and @gmail across the package now returns nothing.

2026-07-23–24 — CJ ENM Mnet Plus Data Engineer application package

Agent: Codex. Target package: career/맞춤이력서/cj-enm-mnet-plus-data-engineer/.

Posting:

Evidence boundary:

Final portfolio order:

  1. infra-k8s-airflow-ops
  2. crawler-modernization
  3. db-index-optimization
  4. binlog-shipping-v4-cdc-upsert

Final validation:

2026-07-23 — ai-frontier study publication: Google ADK 2.5 execution-boundary release

Agent: Hermes cron (06:00 KST) in this repo.

Published one dynamic-fallback AI chapter:

Topic selection rationale:

Control-plane updates:

Verification run required by the study pipeline:

Expected commit message after staging/publishing flow:

2026-07-23 — database-frontier study publication: Materialize v26.33 control-plane contention release

Agent: Hermes cron (05:00 KST) in this repo.

Published one dynamic-fallback database/data-platform chapter:

Topic selection rationale:

Control-plane updates:

Verification run required by the study pipeline:

Expected commit message after staging/publishing flow:

2026-07-22 — AI study publication: AI SDK 7 runtime surface shift

Agent: claude-code. Work folder: /Users/khw/hw/클로드이력서/gs리테일-ai데이터부문/.

Posting (Remember 327461, GS리테일, deadline 2026-07-29, 강남, 경력 3~10년, Remember reward 50만원, process 서류 → AI역량검사 → 1차 → 2차 → 레퍼런스체크 → 채용검진):

하네스 엔지니어링 기반 프로토타입 구축 → 배포·운영, 비개발 직군이 직접 과제를 수정·테스트할 수 있는 환경 설계, 과제 경험의 재사용 가능한 표준화.

직접 설계/운영, Python·SQL, Codex·Claude Code 활용 개발 및 Harness Engineering, 모호한 문제의 구조화.

비개발 직군 교육·코칭, 리테일/이커머스.

Evidence gathering: an Explore agent swept ai/worklog W22–W29, ai/wiki, and portfolio items and confirmed verifiable AI-harness material: the Claude PR-review CI/CD bootstrap kit (secret-scan gate, full-SHA pinning, 5 language profiles), Claude Code/Codex production crawler work across 12 ecosystems (defect taxonomy of 7 types, Java/Maven Python migration, test-verified changes), the Confluence onboarding publication flow (10 read-only code-reader agents, 100-point rubric with a ≥95 gate, 12+12 pages), the LLM license-classification pipeline (SPDX 650+, join-table removal), and the personal 3-agent llm-wiki harness with MCP memory (169 pages synced). Confirmed negatives kept out of the resume: no GCP, no DynamoDB/Lambda, no LangChain/LangGraph, no active vector search (GBrain embeddings are not enabled, so the memory system is never described as RAG).

Deliverables:

  1. 이력서-초안.md — md draft first per workflow; user approved via

"pdf로 변환해줘". New ordering: intro states AI-tool practice and pass-criteria mindset; career bullets lead with the CI/CD kit, crawler modernization, rubric-gated multi-agent documentation, and the LLM license pipeline; new section 개인 AI 시스템 운영 (llm-wiki 3-agent harness, YouTube Shorts platform with 11 live channels); skills add an AI Engineering line at the top. Case 1 = CI/CD kit harness design, Case 2 = 12-ecosystem crawler audit + AI-tool modernization. Known-excluded numbers: Go 99.3% backfill (worklog only, not on a portfolio page), rubric sub-scores, crawler V2 누락률 0% (kept out for consistency with RubyGems 101→4).

  1. 김현욱_GS리테일_AI데이터부문_이력서.html/.pdf — Vroong final HTML

used as the template base, photo base64 carried over; rendered via gstack browse (goto file:///tmp/... + pdf --format a4 --print-background --prefer-css-page-size); 3 pages, each visually verified via pdftoppm.

  1. 김현욱_GS리테일_포트폴리오.pdf — 11 pages (cover + 10 items) via

human/portfolio/selected-portfolio-pdf.html?items=... served on localhost:8765 and printed with browse. Item order: ai-cicd-review-kit, crawler-modernization, ai-license-pipeline-v3, hermes-discord-llm-wiki-system, youtube-shorts-automation-platform, team-doc-governance, aws-to-idc-migration, infra-k8s-airflow-ops, binlog-shipping-v4-cdc-upsert, etl-flow-redesign. All 11 pages visually verified: one item per A4 page, no blank/overflow pages.

  1. 포트폴리오-구성안.md — selection rationale and exclusions

(Forgejo/thesis items, vibekits, personal-portfolio-platform).

  1. 클로드이력서/README.md — added the GS Retail row + posting link;

removed the stale 343GB Elasticsearch doc-count line (fabrication confirmed 2026-07-16; the README still carried it) and left a do-not-reuse note.

Honest-gap strategy recorded in the draft's 메모 section: AWS is real (EC2 only; MySQL and application workloads self-managed on EC2, prod DB still on EC2, 94% cost cut) but GCP, RDS, S3, EFS, and serverless are absent and stay unwritten; interview answers map the DynamoDB/Lambda/S3 requirement to EC2-based infrastructure ownership without implying managed-service experience. AI역량검사 sits between 서류 and 1차 — prep needed if documents pass.

2026-07-22 — AI study publication: AI SDK 7 runtime surface shift

Agent: Hermes cron (04:00 KST) in this repo.

Published one dynamic-fallback AI chapter:

Topic selection rationale:

so the job correctly switched to the 04:00 KST dynamic AI fallback lane.

(2026-06-25) is within the 90-day freshness window, it had no dedicated prior chapter in human/study/content/** or ai/wiki/**, and the release introduces a material runtime-layer change rather than a patch-only feature train.

markdown endpoints (workflow-agent, runtime-and-tool-context, mcp-apps, tool-approvals, file-uploads, skill-uploads, terminal-ui, sandbox reference), plus npm registry publish times for [email protected].

Control-plane updates:

ai/study/curriculum.md with publication date 2026-07-22.

Verification run required by the study pipeline:

Expected commit message after staging/publishing flow:

2026-07-21 — GS Retail resume: 3-persona review loop + portfolio page evidence-chain fixes

Session in Claude Code (/resume skill + Workflow orchestration), work split between ~/hw/클로드이력서/gs리테일-ai데이터부문/ and this repo.

  1. /resume diagnosis of the GS Retail draft: 7-mistake scan graded A

(0 critical). Applied: JD-literal Harness Engineering(...) in the skills line, scope quantification of case 1 result (분석 엔진+크롤러 저장소 전반, user-stated), two bullet↔case dedups.

  1. New resume bullet + portfolio item: 나라장터 공고 선별 시스템

(slug system-associates-infra-portfolio-architecture) — user confirmed 2026-07-21 it is still in real use by 시스템어소시에이츠 영업·마케팅 (비개발 직군). Placed first in 개인 AI 시스템 운영 and inserted as item #4 in the portfolio plan (10→11 pages, user decision).

  1. Persona review loop (Workflow, 5 rounds, CEO/CTO/기술팀장, threshold

95, revision agent constrained to no-fabrication + humanizer voice): scores plateaued 79-83; unanimous verdict 서류 통과(조건부). Blocker was NOT the resume: [상세] links pointed to public pages missing the resume's key claims (evidence-chain breaks). Reviewed working copies were applied over the drafts (backups *.bak-pre-persona-review). Notable resume fixes: unsourced "수십억 행" → "최대 2.6TB", YT mock-first vs live-ops separation, CDC bullet aligned to the v4 page (go-mysql fork/ZSTD details demoted to 메모 as interview material), TimescaleDB dropped from skills.

  1. Portfolio page updates in human/portfolio/items/ (the 7 required

evidence-chain fixes): crawler-modernization ×3 (AI-tool development fact, 누락률 0% reworded as 점검 기준값, code-reader-agents/rubric gate line — grounded in ai/wiki/projects/confluence-lib-onboarding-docs.md); youtube-shorts ("Mock MVP" → platform with mock-first development, live 11-channel operation split into its own paragraph); system-associates (실사용 중 2026.07 기준 in 핵심 요약); ai-license-pipeline-v3 ×2 (unmeasured "더 깊고 정확한" wording → verifiable structure facts); binlog-shipping-v4 (legacy-inclusive ~10 on-premise customers). check_human_voice.py: BAN 0.

  1. Resume HTML regenerated from the revised md (same Vroong-base

layout, photo kept); PDFs re-rendered after this commit — see the application folder. Remaining open items in the draft's 메모: LangGraph 학습 중 표기, 병역사항, 연구소 기관명 표기 통일.

2026-07-22 — GS Retail AWS service-scope correction

User-confirmed source: ai/sources/career/2026-07-22-aws-service-scope-correction.md.

were self-managed on EC2.

and slide wording. RDB is treated as the MySQL workload, not as proof of RDS.

on AWS while other EC2 workloads moved to IDC/In-house, cutting monthly AWS cost by about 94%.

2026-07-22 — GS Retail review corrections and spec-driven validation evidence

User-confirmed source: ai/sources/career/2026-07-22-spec-driven-crawler-review-clarification.md.

keeping the evidence boundary explicit: no verified repo count, issue-catch rate, or time-reduction metric.

crawler change by another owner could alter the produced data format; the current workflow uses spec.md as the contract and checks implementation diffs through profile tests and AI review.

with each owner; Hyunwook owns cross-ecosystem audit, common data-model and execution concerns, selected crawler modernization, and the Java/Maven Python migration.

replaced by an internally owned pipeline, but the ChatGPT API remains an explicit model dependency.

transformed DML. Truncation skip is now stated as a separate reconciliation boundary rather than as lossless processing.

2026-07-22 — Shared job-tailored resume quality pipeline

User source: ai/sources/career/2026-07-22-cross-agent-job-tailored-resume-pipeline-request.md.

Code and Codex use the same job-posting capture, requirement map, evidence selection, canonical template, review, and artifact-validation flow.

requires score >=95, independent CEO/CTO/tech-lead scores >=95 each, and zero hard-gate violations for unsupported claims, contradictions, role inflation, placeholders, broken links, public evidence mismatches, or PDF defects.

requirements remain honest gaps and cannot be converted into claimed production experience to increase the score.

career/맞춤이력서/<company>-<role>/ and the required per-application artifacts (job-posting.md, requirement-map.md, sources, HTML/PDF, review-manifest.json, and quality-gate.md).

Codex and imported by CLAUDE.md.

unsupported claim, reviewer below 95, inconsistent counts, and source-level placeholder detection). Python syntax, JSON, YAML/frontmatter constraints, git diff --check, and human voice checks pass; human voice remains BAN 0 with two unrelated pre-existing warnings.

Resume AI-voice ownership wording

“with Claude Code·Codex.” The resume now states that the candidate set the problem, change scope, output contract, and validation criteria; the tools produced drafts or patches; and only changes passing tests and data checks were applied.

folders. PDF regenerated as a three-page A4 document.

existing WARN remains for the generic word “현대화” in the resume case title.

is now “누락 조사와 Python 전환 검증” in the active HTML/PDF artifacts. Backup files retain historical wording and are intentionally not used as submissions.

점검과 Python 전환” and passes BAN 0 / STRUCT 0 / WARN 0. A separate legacy GS Retail portfolio PDF still contains the old title; it is unlinked and must be regenerated from the current portfolio source before submission.

2026-07-22 — GS Retail portfolio PDF renderer restoration and visual QA

worklog and the pre-regression implementation: serve selected-portfolio-pdf.html?items=items/<slug>.html,... through localhost:8765, wait for network idle, and print with the gstack browser using A4, CSS page size, and background graphics. Bare slugs remain invalid.

global page scaling. Restored content-aware fitting and added a one-pixel tolerance for Chromium's A4 rounding so an exact 297mm page is not needlessly reduced to 98%.

only duplicated 배경/역할 sections already represented in the header or linked web detail page. Source detail pages remain intact. Runtime inspection confirmed zoom: 1 for the cover and every item page.

with flex sizing and increasing inter-card/callout spacing. This prevents the Hermes curation column from extending through the highlighted note.

pages inspected as PNG renders; no clipping, overlap, blank spill page, or inconsistent type scale; all 11 public detail URLs extracted from their expected PDF pages; PDF structure check passed; strict human-voice check reports BAN 0 / STRUCT 0. The canonical Drive copy and local compatibility copy have SHA-256 69d349ba7023b47df79fb41007c72e6a26b264e250e3c808e42884ab7cd33b5a.

assembled PDF: four team-governance prefix labels inherited the browser's default 16px size, and one CDC architecture paragraph retained an inline 13px size. Added PDF normalization for both; a computed-style audit now finds no body leaf above the 12.2px section-title threshold.

decision sentences use a blue callout, verified figures and key constraints use a marker underline, result summaries remain metric cards, and caveats use an amber callout. The four team-governance work types now render as consistent blue chips. The ETL 역할 section is omitted only from the one-page PDF summary because ownership is already present in the header; its web detail page remains unchanged.

Every page remains at zoom: 1; A4 structure and all 11 detail URLs pass. The current Drive and local copies have SHA-256 be6c0159c61912c5e264316f4bb55df87b5fd90ca4cab8c24a571c41b3521cdb.

full-card internal link to its item page and shows its destination page number. Each item footer now includes a return link to the cover contents.

return-to-#1 link on each of the 11 item pages. The 12-page A4 layout and zoom: 1 type scale remain unchanged. The linked Drive and local copies have SHA-256 0ae72c79d1fa8bf698389456a938eb48d2e78186f09645af6c6a95ea20e9b9e0.

card links as named destinations (LINK_NAMED) resolving to destination point (0, 841.92), the bottom of each A4 page. Link extraction alone had therefore overstated viewer compatibility.

rendered PDF. It retains each full-card annotation rectangle, replaces named destinations with direct PDF /GoTo actions, and sets every target to the top-left (0, 0). It also normalizes the 11 return-to-cover links and fails if the expected counts or destination coordinates do not match.

verification reports 11 cover rectangles targeting pages 2-12 at (0, 0), 11 return links targeting page 1 at (0, 0), and no named internal links. The current Drive and local copies have SHA-256 60370e8d5888e96d520243f722e507973f29301b885ab904467c3715dca62315.

2026-07-22 — GS Retail evidence and human-voice finalization

ai/sources/career/2026-07-22-gs-retail-ai-data-project-role.md and mapped every requirement to verified evidence or an explicit fit gap.

acceptance rule: accuracy 35, completeness 25, diagram quality 20, format compliance 10, and readability 10; code spot-checks are required and a document below 95 is rewritten. This wording appears once in the resume and is backed by the detailed portfolio page.

copy separates problem definition, output contracts, tests, data checks, Claude spec.md review, and Bitbucket merge responsibility instead of attributing the work broadly to Claude Code or Codex.

Discord alerts and follow-up questions used by sales and marketing, while an operator adjusts separated settings; a non-developer self-service editing UI remains an explicit gap. The MySQL page no longer invents a primary-to-replica rollout order, and the YouTube page separates mock-first tests from the confirmed 11-channel operating scope.

five posting-direct items and six operating-capability items. Each item stays on one A4 page with a consistent type scale and restrained highlight system.

pages were visually inspected with no clipping, overlap, blank spill pages, or broken glyphs. The portfolio contains 11 cover links to pages 2-12, 11 return-to-cover links, and 11 external detail links. The current SHA-256 values are 387f828f6f45d98989047a8dfd430ab592c3950df07a4deef49c65f016b618e2 for the resume and 9f1c967ceda9777b4e1541264116f71252e8a2547b715c94a1bad3124eea3924 for the portfolio.

blocking issues. The deterministic gate passed at 95.3/100. This remains a document-quality and evidence-discipline score, not an estimate of hiring probability. Honest gaps include GCP, managed AWS services, LLM orchestration frameworks, RAG/Vector Search, retail domain experience, and a direct non-developer self-service UI.

0; scorer tests 6/6 and voice-check tests 5/5 pass; Python compilation and git diff --check pass. The relative-link checker was corrected to ignore runtime template placeholders and strip cache query strings before resolving local paths, then passed against the full portfolio tree.

운영 역량 with the reader-facing 먼저 볼 작업 and 함께 볼 작업. The regenerated portfolio remains 12 A4 pages at zoom 1 with 11 cover links, 11 return links, and no clipping. The previous PDF is preserved as 김현욱_GS리테일_포트폴리오.bak-20260722-2055-before-toc-labels.pdf.

production DB index case while keeping the public YouTube page. The new item follows the AWS-to-IDC case and records the verified results: collection DB 9.6TB to 4TB, distribution DB 2.6TB to 1.2TB, and customer on-premise DB creation 12 hours to 6 hours. Its generic 작업자의 메모 heading was replaced with the decision-specific 인덱스를 바로 지우지 않은 이유. The regenerated artifact remains 12 A4 pages at zoom 1 with 11 cover links, 11 return links, and 11 external detail links; its SHA-256 is bcf6104c3a874be7a2b659eda8459bd9d0933df2072322a597a09f4570864827. The preceding submission PDF is preserved as 김현욱_GS리테일_포트폴리오.bak-20260722-2106-before-db-index-swap.pdf.

2026-07-22 — 당근 Software Engineer Data 지원 패키지 + humanizer 파이프라인

  1. 당근(데이터 가치화팀) 이력서·포트폴리오 구성안 작성, 3-persona 리뷰

루프 10라운드 (80→94 수렴, 라운드 9 CTO 97). 최종 94/94/94 정직 보고 — 상세는 Drive career/클로드이력서/당근-software-engineer-data/quality-gate.md.

  1. 사용자 지시로 humanizer("i'm not ai") 패스를 파이프라인 필수 단계로 추가:

skills/tailor-resume-to-job/references/humanizer-pass.md (blader/humanizer 33패턴의 한국어 이력서 적용판) + SKILL.md 4b단계 + 3-persona 리뷰의 AI 티 검사(ai_tell_findings) 의무화.

  1. 리뷰 루프가 지적한 포트폴리오 페이지 AI 티 43+곳 재작성 (사실 무변경):

em dash·따옴표 강조·아포리즘·부정 대조 클러스터·명사형 종결·메타 문장· "~한 이유" 헤더 공식 4곳 교체. check_human_voice.py BAN 0/STRUCT 0.

  1. AWS RDS 표기 정정(사용자 확인: EC2만 사용) — aws 페이지 태그에서 RDS 제거.
  2. NHN Cloud 보안 AI엔지니어 초안·구성안 작성 (보안 도메인 전면, ML 정직 갭

명시). 병역 확인 대기 — NHN은 병역 필·면제만 지원 가능.

2026-07-22 — GS Retail final evidence correction (supersedes earlier finalization entries)

series (8.0.29) from MySQL 8.4.0 LTS instead of describing both as LTS series.

ai/wiki/projects/bts.md: Go-based offline ROW event parsing, MySQL 8.4 ZSTD payload handling, and duplicate-safe UPSERT re-execution for normally converted INSERT and UPDATE events. Removed unsupported checkpoint, hash, smart-polling, automatic-rollback, and direct-production-deployment claims. The submission now explicitly excludes DELETE, DDL, and truncation skip from the re-execution guarantee and separates the user's implementation and support role from the technical-support team's customer deployment and daily operations.

and public detail page. Added a --slug option to the diagram generator so a single changed diagram can be regenerated without touching unrelated pages. Published portfolio input commit: 8d448b9; all 11 selected local detail pages exactly match their deployed copies.

portfolio (12 A4 pages). No clipping, overlap, inconsistent body type scale, or broken glyphs was found. The portfolio has 11 cover-to-item and 11 item-to-cover direct LINK_GOTO annotations at (0, 0), no LINK_NAMED annotations, and 11 valid external detail links. All 16 unique external URLs in the resume return HTTP 200.

299b5f5eceeb80c7214869bdb92218bb7bbe803195883653dd4e2e12d438f3c7; portfolio 78e9d3470f5f967facd8f023c6d656d504b5542c3b8a5fba8f5d4b3d0db30233. The Drive and local compatibility copies match byte for byte.

deterministic document-quality gate passes at 95.3/100 with no hard-gate failures. Evidence coverage is 40/40 resume claim units, 43/43 numeric claims, and 11/11 portfolio evidence groups. This score evaluates document quality and evidence discipline, not hiring probability.

scorer tests 6/6; human-voice tests 6/6; diagram generator compiles; all 15 output pages were visually reviewed.

2026-07-23 — NHN Cloud 보안 AI엔지니어 application package (데이터보안분석팀)

경력/정규/판교 삼평동). Resumed a paused draft (이력서-초안.md + 포트폴리오-구성안.md). Work folder career/클로드이력서/nhn-cloud-보안-ai엔지니어/.

(2010.08~2012.08) = 병역필, so the posting's male 병역필/면제 gate is met; and the undergraduate degree is 숭실대학교 정보보호학과 학사 (2016.03~2018.08, GPA 3.96/4.5). Both undergrad and master's are in 정보보호학 — a security-domain double track for this posting. Added a 병역 line and the undergrad row (the posting asks for full education history; the four earlier resumes listed only the master's).

new top skill groups Security Data + AI Engineering, career bullets and cases reordered so 라이선스 LLM 파이프라인 and 전수 대조 자동 점검 read first. Rendered a 3-page A4 PDF via headless Chrome and visually verified every page.

selected-portfolio-pdf.html. The sejong-master-thesis item overflowed one A4 page (footer spilled), so added it to pdfSectionOmissions (drop 작업자의 메모 + 의의, keep 문제·접근·결과·성과·원본 자료) — every item is now exactly one page.

"서류 통과 가능(조건부)". Applied every actionable finding: bullet re-order, 소개 담백화, 사례1 결과 숫자 마감, 사례 제목 em dash→콜론, RubyGems 수치 층위 분리, 사례2 표준화 오너십 완화.

학력, 포트폴리오 필수, 미검증 스택 0, human-voice BAN0/STRUCT0, 링크 14/14 200, PDF 시각 검증). The 95-unanimous document-quality gate is NOT certified — not a document defect but an inherent DE→보안 AI엔지니어 stretch (자격요건1 5년 ML 모델 운영 미충족). Did not invent ML experience to lift the score. Weakest interview point recorded: LLM 라이선스 분류(650건)의 정확도 검증 근거 부재.

BAN 0 / STRUCT 0 / WARN 2 (다양한, 허용); 14 external resume links all HTTP 200.

omission entry) and ai/sources/career/2026-07-23-nhn-cloud-security-ai-engineer-posting.md.

2026-07-23 — 코오롱베니트 데이터 엔지니어 application package (headhunter scout)

데이터 엔지니어, 정규직, 과천. Work folder career/클로드이력서/코오롱베니트-data-engineer/.

posting's required skills (Spark/PySpark + OOM tuning, Delta Lake/Iceberg) and core duties (Databricks/Cloudera lakehouse, real-time streaming) are all unheld. Per the fixed rule, none of Spark/Delta/Iceberg/Databricks/streaming/ lakehouse terms appear in the resume. The winning axes are the two 자격요건 (DE 5년 + led a platform design/build) and technical leadership.

the 80+ crawler batch platform and 팀 개발 체계, case 1 is the AWS→IDC migration with the 3-way DB split, case 2 is the RAW-preserve ELT redesign. Resume 3 A4 pages, portfolio 11 pages (cover + 10 platform/leadership/large-scale items).

full banned-stack scan and found zero occurrences. Applied every actionable finding: headline scope+leadership signal, case-1 result de-hedged ("이관에 직접 기인한 장애는 없었고 초기 안정화 조정 몇 건" — matches the verified ~2-issue fact), case-2 measured-vs-estimated boundary made explicit, title em dash removed.

stack 0 via CTO scan, human-voice BAN0/STRUCT0/WARN0, links 14/14 200, PDF visual check). The 95-unanimous gate is not certified — an inherent stack-fit gap, not a document defect; did not invent Spark/lakehouse experience.

ask whether the client weights stack or platform-lead) but per user direction proceeded straight to building the resume/portfolio package.

2026-07-23 — Toss Place DAE document-screen rejection retrospective

document-screen rejection. Canonical status is rejected; highest confirmed stage is applied. A supplied email screenshot shows the rejection notification at 2026-07-22 14:15; the exact application date remains Unknown.

ai/sources/career/2026-07-22-toss-place-dae-rejection-email.png. The email says incumbent-role staff reviewed the documents and frames the decision as choosing someone more suited to the Data Analytics Engineer role.

ai/sources/career/2026-07-23-toss-place-dae-application-result.md and updated ai/wiki/projects/2026-career-transition.md.

extraction succeeded; no clipping, overlap, blank page, or broken glyph was found. The resume human-voice check returned BAN 0 / STRUCT 0 / WARN 0.

on Snowflake, dbt, formal DW modeling, SSOT, marts, and product metrics, while the package's strongest evidence was MySQL schema/query work, data collection, reconciliation, Airflow/Kubernetes operations, and infrastructure.

transferable SQL, data-quality, and standardization evidence too much like direct Analytics Engineering and DW experience.

explanation is recorded as official wording, while the Snowflake, dbt, DW, and product-metric gap analysis remains explicitly labeled as inference. No resume or portfolio artifact was modified.

2026-07-23 — NHN PAYCO 데이터 엔지니어 application package

build the NHN PAYCO 데이터 엔지니어 package instead (posting was already preserved at ai/sources/career/2026-07-23-nhn-payco-data-engineer-posting.md by another session). Work folder career/클로드이력서/nhn-payco-data-engineer/.

OLAP DW/마트·큐브 are unheld, but unlike 코오롱 the required axes lead with ETL·데이터 파이프라인 (자격1) and SQL·데이터 모델 설계/튜닝 (자격4), which are direct strengths, so the winning surface is wider. None of Hadoop/Spark/Trino/Hive/Kafka/Flume/Druid/ ClickHouse/OLAP-cube appear in the resume; large scale is described as MySQL-based.

파이프라인 + DB 모델·인덱스 튜닝, case 1 is the RAW-preserve ELT redesign, case 2 is the EXPLAIN-driven index/schema redesign (95 queries, 224 indexes/72%, 2.6TB→1.2TB, 9.6TB→4TB). Added the 병역 line (posting gates on it). Resume 3 A4 pages, portfolio 11 pages (cover + 10 ETL/SQL-tuning/standardization items).

banned-stack scan → zero occurrences. Applied every actionable finding: case-2 result split so each action maps 1:1 to its TB reduction, crawler cell retitled for ownership clarity ("12개 생태계 누락 점검 표준화(8개 전수 대조)") and "12개 언어 생태계" wording to avoid confusion with "80+ crawlers". The index-72% judgment basis (95 실사용 쿼리 EXPLAIN) already exists on the db-index-optimization page — consistent, nothing invented.

병역·학력 게이트, human-voice BAN0/STRUCT0/WARN0, links 13/13 200, PDF visual check). 95-unanimous not certified — inherent Hadoop/Spark/OLAP stretch, not a document defect.

2026-07-23 — NHN PAYCO Data Engineer final release (supersedes earlier entry)

posting and released a distinct package under career/맞춤이력서/nhn-payco-data-engineer/. This entry supersedes the earlier draft-package review above; the application remains not submitted.

pipeline work, Python, SQL/EXPLAIN, MySQL schema/index tuning, Linux scripting, data standardization, and platform automation. Kept Hadoop/Spark/Trino/Hive, formal OLAP mart/cube delivery, Kafka/Flume, analytical OLAP engines, Hadoop operations, and PAYCO domain work as explicit gaps.

ambiguous portfolio selections. The final ten-item portfolio substitutes the source-backed AWS-to-IDC migration and 767GB backup-automation cases for pages whose estimates or unmeasured outcomes could be read too strongly.

index optimization, and AWS migration. Verified all ten selected public pages are byte-for-byte identical to the local sources used for the submission PDF.

bounds exposed a possible overlap. The final page-4 label ends at y=64.2375 and the title begins at y=66.9000; all eleven portfolio pages were visually checked again. Internal links were normalized to ten cover-to-item and ten item-to-cover GoTo annotations.

The deterministic document-quality gate passed at 95.6/100; must-have evidence coverage is 80% and preferred evidence coverage is 50%. This is a document quality result, not a screening-probability estimate.

PDF pages visually inspected; human-voice check BAN 0 / STRUCT 0 / WARN 0 across the resume and ten selected portfolio sources; scorer tests 6/6; no unsupported claims, role-boundary violations, production-stack inflation, placeholders, clipping, broken links, or public-source mismatches.

42453efd978c390feb24563f1df1d9631c63e657d97d9f1dd1aec26c719f4532; resume PDF SHA-256 3aaf9608235ba20eb4522504e4acf8f17183e2238ec5040bd3e30c5e82a3fc04; portfolio PDF SHA-256 e672fbd7a1e138aacd2a0afb5d99bc684b882a043e7125bfecdc53a9a287c462.

2026-07-24 — NHN PAYCO resume richness and ATS revision

asked to use the existing Claude-generated NHN, Carrot, and GS Retail resumes as density references. Compared those variants with the evidence-gated final package; retained their useful operating detail but rejected unverified AWS S3/EFS claims, unrelated AI material, and unsupported big-data production stacks.

pages. Added source-backed detail for per-crawler DAG/Pod resource control, Git-Sync, SeaweedFS RAW preservation and version_id, 12-ecosystem data-quality checks, DML Broker backpressure/idempotency, the batch CDC role boundary, incident inspection order, AWS-to-IDC migration, and collation/schema repair.

that a two-column operations grid produced incorrect PDF extraction order. Removed repeated summaries, moved the collation case to the operations section, replaced the stale DB-engine comparison card, and changed the five operations entries to a single-column DOM. pdftotext now returns each title immediately followed by its body in CDC → monitoring → backup → collation → AWS order.

FIRM-COV is the follow-up IEEE Access paper in which the user is third author. Preserved the previously recorded user-confirmed education, GPA, and military facts in ai/sources/career/2026-07-23-education-military-user-confirmation.md and updated the canonical person page and resume template to use those boundaries.

overlap, blank page, or broken glyph; canonical template CSS unchanged; human-voice BAN 0 / STRUCT 0 / WARN 0; 14 linked-claim occurrences and ten unique portfolio pages; all ten public caches and deployed pages match the current local sources; no placeholders or prohibited production-stack claims.

deterministic document-quality score remains 95.6/100, with 80% must-have and 50% preferred evidence coverage. The direct Hadoop/Spark/Trino/Hive, OLAP mart/cube, Kafka/Flume, analytical-engine, Hadoop-operations, and PAYCO-domain gaps remain.

69a34ed7cff149c27fa6babf29104e21ba8cd31ceec95de6656d43bccac83787; PDF 8bf621ef509a514df1090228c3281532ad1ab23721f824c25e54c4e454a6433f. The portfolio PDF was not changed.

2026-07-24 — NHN PAYCO LabradorLabs career expansion (supersedes the richness layout above)

LabradorLabs, not the resume's total text volume. Moved the five separately grouped operations items into the employer history and removed the duplicate later section.

three under an explicit 경력 (계속) heading on page 2. The sequence covers Airflow batch operation, RAW reprocessing, source quality checks, EXPLAIN/index work, DML backpressure, batch CDC, collation/schema repair, monitoring, 767GB backup automation, and the AWS EC2-to-IDC migration.

P6 to item 6. Independent rereview passed at CEO 98, CTO 99, and tech lead 98 with no blockers.

unchanged; fourteen linked-claim occurrences across ten unique public detail pages; all three pages visually clean; human-voice BAN 0 / STRUCT 0 / WARN 0; no placeholders or prohibited production-stack claims.

bca6e73697d60e9fff2dbb9d07bbbb1319ae8d6c039a1f60ced88071c7259f89; HTML 49dccff5cdaf065acb3502f7a8af8396c908573587bd41a2559b13b5b19cd6a7; PDF bad5fd437f6f02bf6fef065a26d244c1083729e7974aa0a58038bf560fb26de2. The portfolio PDF was not changed.

2026-07-24 — Application-status snapshot and 코오롱베니트 no-go decision

User-reported status review (2026-07-24). Recording it here so future sessions do not re-ask or re-derive submission state.

Company / roleChannelSubmittedResult
리멤버앤컴퍼니 Data EngineerRememberYes서류 통과 (document pass)
부릉(Vroong) Data EngineerRememberYespending
토스플레이스 DAETossYes불합격 (document-screen reject, 07-22)
토스페이먼츠 Data EngineerTossNo — skipping this round—
GS리테일 AI데이터부문RememberPlanned today (deadline 07-29)—
CJ ENM Mnet Plus Data EngineerWantedBuilding + submitting today—
NHN PAYCO Data EngineerNHN careersPlanned today—
당근 Software Engineer, DataDaangn careersPlanned today—
NHN Cloud 보안 AI엔지니어NHN careersNo — declinedtoo far from current work
코오롱베니트 Data EngineerHeadhunter (써치라인)No — declinedsee reasoning below

Confirmed changes vs prior records: 부릉 was submitted (previously "확인 필요"); 토스페이먼츠 is intentionally skipped this round; 리멤버앤컴퍼니 is the source of the one 서류 통과. GS/CJ/NHN PAYCO/당근 are the four the user submits today.

코오롱베니트 — decided NOT to apply (2026-07-24). Two stacked mismatches, and the user is not in a position that requires a stretch (already one document pass at 리멤버 plus four better-fit applications going out today):

  1. Stack gap: the posting's required skills (Spark/PySpark + OOM tuning, Delta

Lake/Iceberg) and core duties (Databricks/Cloudera lakehouse, real-time streaming) are entirely unheld. This is the same "core stack absent, only transferable SQL/schema evidence" pattern that produced the 토스플레이스 document-screen rejection — and the gap here is wider (there the missing stack was dbt/Snowflake as requirements; here Spark/Databricks are the premise of the main duties). So screening pass probability was judged low.

  1. Direction mismatch (the deciding factor for the user): 코오롱베니트 is an SI

(systems-integration) shop. The user's career to date is deep single-product data-platform ownership (design → build → multi-year operation of one system: CDC, ETL, DB tuning). SI shifts the center of gravity to broad multi-client project delivery plus PM/PL, i.e. a directional change away from the user's existing IC depth rather than a continuation of it.

The clean exit line: decline via the headhunter as "이번 포지션은 방향이 맞지 않아 고사" — keeps the relationship and lets them surface a better-fit posting later. A headhunter reply draft exists at career/클로드이력서/코오롱베니트-data-engineer/헤드헌터-회신-초안.md (disclose the Spark/Databricks gap up front) but the user chose to decline outright.

Career axis — DECIDED (2026-07-24): deepen as an IC. The user prefers to keep going deeper on single-system/product data-platform work rather than moving to a leading/PM track. Future target selection: prioritize product/platform IC roles; weight down pure SI or PM/lead-centric postings (this is the basis for the 코오롱 no-go). Leading experience is still usable as a strength, but the career goal is IC depth.

Durable lesson reinforced (from the 토스플레이스 result): for postings whose core required stack is absent, transferable SQL/Airflow/schema experience does not make them high-fit. Treat them as stretch and, on a headhunter channel, confirm the client's real priority before spending a submission.

2026-07-24 — Resume highlighter standard + application-status hub page

auto-renders as a yellow marker via CSS in the canonical template (ai/templates/resume-ats-template.html): b{background:linear-gradient( transparent 56%,#ffe27a 56%)} with .contacts b, .skills b, .job-intro b excluded so labels and job titles stay unmarked. Documented it in AGENTS.md "Resume Assets" as a binding cross-agent rule (Claude/Codex/Gemini all use plain <b> from the template → identical highlighter nuance; no separate mark style). Applied the same rule to all four today-batch resumes (GS리테일, NHN PAYCO, 당근, CJ ENM); each still renders 3 A4 pages.

gate 95.6, with a grounded "분석 모델링 학습" DW-modeling line) is the submission version. Copied it into career/클로드이력서/nhn-payco-data-engineer/ (fixed the 학과명 to 정보보호학과 in two places) with the Claude-built version moved to _deprecated-claude/.

human/career/index.html (styled with the hub's wiki.css, all applications in tables with URLs/scores/status/folders), linked from human/index.html nav + card, with a /career redirect in human/_redirects. The hub is Cloudflare-Access-gated (owner-only), so the job-search status and rejection data are not publicly exposed. Source markdown: human/career/2026-지원-현황.md.

2026-07-24 — Four planned applications submitted + NHN PAYCO site-form record

The user confirmed that all four applications in the 2026-07-24 batch were actually submitted. This supersedes the earlier same-day snapshot that marked them as planned or in progress; that earlier table remains as a point-in-time record.

Company / roleSubmission channelFinal status
GS Retail AI Data DivisionRememberApplied (2026-07-24)
NHN PAYCO Data EngineerNHN CareersApplied (2026-07-24)
Daangn Software Engineer, DataDaangn CareersApplied (2026-07-24)
CJ ENM Mnet Plus Data EngineerWantedApplied (2026-07-24)

NHN PAYCO submission method and retained record:

김현욱_NHN페이코_포트폴리오.pdf (1.8 MB). The portfolio URL entered in the form was https://portfolio.hwlabs.dev/.

fourteen skills, and the final 886-character cover letter were entered directly in the NHN Careers application form.

lead 97, with zero blockers and human-voice BAN 0 / STRUCT 0 / WARN 0.

Trino, Hive, Kafka, Flume, Druid, ClickHouse, formal OLAP marts/cubes, or the PAYCO payments domain.

method is retained at human/career/최종 NHN PAYCO 공고 사이트.pdf.

2026-07-27 — Remember DuckDB submission one-take packaging

running Docker Engine and Compose v2; it resets generated state, builds and starts PostgreSQL/Airflow, waits for DAG registration and 2022/2023 completion, prints task logs and database inspection output, and exports the completed DuckDB into the extracted directory.

1. Docker preserved Google Drive's 0600 mode on requirements.txt, then changed ownership to root, so the Airflow image user could not read it. 2. Docker Desktop bind mounts over the Google Drive File Provider path repeatedly returned errno 35 (EDEADLK) during Python imports. 3. macOS NFD Korean CSV filenames did not match the pipeline's NFC filename template on Linux.

code/data, NFC-normalized image filenames, a DuckDB named volume, and an atomic host export after success. No .env, host Python, host DuckDB CLI, or host bind mount is required.

SUCCESS; 30,513 initial rows; 2023 inserted 2,174, updated 4,000, unchanged 17,087, deactivated 9,426; master 32,687 with 23,261 active; assignment-rule mismatches 0; Airflow import errors none; container tests 108 passed in 2.22s.

Archive validation: 52 entries, 1,667,413 bytes, all UTF-8 and NFC paths, executable up.sh mode 0755, SHA-256 54b6c94210baa2ec347c0ea04ef35cf9200becfd59b74306e8056c100afe89e5.

2026-07-27 — Remember final technical-lead and AI-slop review

technical judgment, code naturalness, human voice, abstraction discipline, evidence consistency, and reproducibility. The first pass scored 54/100.

empty files, and database-inspection errors. Added regression tests that first reproduced each failure.

checks and query errors now roll back the master and change history before the sync run is closed as FAILED.

merge.py to a 41-line public facade. The Airflow DAG now has four typed tasks and one policy snapshot per run.

generic claims from README, architecture notes, reports, tests, and comments. A Korean AI-phrase scan found no marketing clichés or AI/Claude/Codex attribution in the submission.

clean; BasedPyright 0 errors, 0 warnings; Compose config and Bash syntax valid; Python no-excuse audit clean across 35 files.

could not start because their access token refresh failed, so their results were not counted or inferred.

the DuckDB choice and its single-writer tradeoff, and the ./up.sh execution and database-inspection path in concise Korean. A separate independent technical-lead pass scored it 91/100 at first, then 99/100 after replacing unsupported EXCEPT and anti-join claims with the actual read_csv, IS DISTINCT FROM, LEFT JOIN ... IS NULL, and FILTER usage, narrowing update/WARN wording to the implementation, and removing formulaic phrases. This was a documentation-only archive change; CRC, source parity, UTF-8/NFC paths, executable modes, and DB/cache exclusion were revalidated.

document: judgment in 과제 작업 설명.md, operations in README, source evidence in reports/data-profiling.md, run evidence in reports/run-result.md, and system structure in ARCHITECTURE. README dropped from 258 to 119 lines and the work explanation from 113 to 54. Replaced the ASCII architecture with three Mermaid diagrams for pipeline flow, Docker runtime, and the logical data model. Mermaid CLI 11.15.0 rendered all three, relative-link and fence checks passed, and the narrative documents returned BAN 0 / STRUCT 0 / WARN 0 from the human-voice check.

edits, using the gstack plan-eng-review rubric (severity plus confidence, and the rule that a finding needs a quoted source line). Score 84/100 against Codex's self-scored 96/100. Six items fixed.

three deliverables and names README.md as the place for 설계 시 고려한 중점 사항. The document-boundary cleanup had moved all design rationale into 과제 작업 설명.md, leaving README purely operational. README now carries nine numbered decisions with the observed numbers behind each: key handling without padding, IS DISTINCT FROM change detection after normalization, the blank-2023 policy and its option, soft deactivation of the 9,426 absent workplaces, the RAW/STAGING/MASTER split with a column-level change log instead of SCD Type 2, the single-transaction apply with eight reconciliation checks, rerun safety with the year-regression guard, and the DuckDB choice with its single-writer tradeoff. 과제 작업 설명.md became a cover note mapping the three deliverables to files.

합니다체 and continued in 하다체; data-profiling.md was 하다체 with a 합니다체 closing paragraph, and the split fell exactly on the paragraphs the cleanup had appended. Each document is now one register. Three documents also ended with the same three-way link enumeration; rewritten or dropped. Cross-document duplicate lines over 30 characters went from 1 to 0, banned-vocabulary hits stayed at 0.

suppressing nothing (BLE001 is real) in a package that ships no Ruff config. Replaced with the actual reason for the broad catch.

three transactions while the DAG and CLI both call apply_master, which runs them in one transaction with rollback. The helper now calls apply_master; the two production-dead wrappers and the __all__ hasattr test that pinned them were deleted. Twelve int(str(scalar(...))) sites collapsed into db.scalar_int().

up.sh to 0600, so ./up.sh failed with permission denied in the working copy. The old ZIP still held 0755 from an earlier copy, so a rebuild would have shipped a package the reviewer could not start. The recipe in HANDOFF.md now re-applies 0755/0644 in the staging copy and asserts both.

CLI run on the real CSVs reproduced 30,513 / 2,174 / 4,000 / 17,087 / 9,426 and master 32,687 with 8/8 reconciliation on both years; ./up.sh in Docker exited 0. The rebuilt ZIP (52 entries, 1,669,240 bytes, SHA-256 04c1c405…8713e) was extracted with /usr/bin/unzip into a Korean-named directory, where ./up.sh again completed with both runs SUCCESS, source/master mismatches 0, and 107 container tests passing. Post-fix score 96/100.

the deliverable map, the design decisions, and the run steps now live in one file (165 lines) instead of being split across two. Dropped the duplicated "현재 범위" block from ARCHITECTURE so the limits are stated once. Cross-document duplicate lines stayed at 0, each document keeps one Korean register, links resolve, and the banned-vocabulary scan stays clean. Rebuilt ZIP: 51 entries, 1,667,277 bytes, SHA-256 43adf7f5…3426a, with local tests at 101 passed / 1 skipped and zip↔working copy byte parity confirmed.

be a MotherDuck-hosted app: it downloads the ui extension at runtime, binds only to the container's IPv6 loopback (so a socat relay is needed in Docker), and its frontend failed with "Initialization Error - Failed to resolve app state with user" on two separate origins. A reviewer meeting that screen would be worse than no UI.

copy-pasteable read-only duckdb.connect(...) query that needs no local DuckDB, the six table names, a pointer to sql/verification_queries.sql, and the test command. The paste was executed verbatim to confirm it runs. README's verification section matches. Re-ran ./up.sh end to end (exit 0, master 32,687 / active 23,261, mismatches 0) and rebuilt the ZIP: 51 entries, 1,668,229 bytes, SHA-256 7caa1323…ff85.

opinion: every table is read by multiple call sites (master 22, staging 10, change log 8, raw 6, quarantine 6, sync run 5), so none is decorative. Kept the schema — removing one would cascade through the DDL, the eight reconciliation checks, 107 tests, four documents, and the ERD on the day before the deadline. Instead README now states plainly that the brief requires only master_workplace and lists what each of the other five is for in one sentence each. ARCHITECTURE's storage-layer table keeps the rerun semantics and no longer repeats the rationale.