01
Fifteen years of data
20,010 original files · 548.5GB. Scale across 11 themes and the year-by-year flow.
Year × theme (2011–2026)
Stacked bars · colors match the themes above
Document types
File types (top 12)
02
Problems in a scattered archive
Duplicates, draft versions, scanned PDFs, and HWP files block search and cleanup.
03
Pipeline into knowledge
Ten steps from a read-only copy to search and a link table. Scroll lights each step and its measured numbers.
04
How the AI worked
Hands-on demos with measured and example data. No source text or real filenames.
② Meaning map
③ Hybrid search
④ Read and extract
⑤ Group drafts
05
How accurate
Human spot checks. Strong results and weak spots together.
-
99.8%
Text extraction
12,519 / 12,549 · 21 HWP failures after retry
-
30/30
Version grouping
30 random groups checked by a person · after rule fixes
-
8/10 → 10/10
Document labeling
First pass 8/10 (2 year errors) · extension set 10/10
-
6/10
Link table · still weak
Project/org extract · 2 errors were non-project docs mixed in
Semantic search by query (on-theme in top 10)
A person opened the top 10 for each query and judged them
- Factory fire prediction10
- Public-client BIM training9
- Underground disaster8
- Biosignal danger detection4–5
06
Principles
Data stays local. A person makes the last call.
-
Read-only originals
Originals are copied, never deleted or edited. Checked with sha256 fingerprints.
-
Local models
Labeling, summaries, embeddings, and OCR run on the office server. Nothing leaves; API cost is 0.
-
Auto detect & mask sensitive data
Finds resident-ID, account, and password patterns plus certificate files, logs locations only, and masks some search body text with ●●.
-
Private-network search UI
The search UI opens only on the private network. This public page has no full-text search.
-
Human final check
Spot checks and rule fixes are human. AI output is not used as a record as-is.
07
How it is used now
Find files · project history · paper leads · performance lists. Below is a fictional mock screen.
Example screen
Search · “tunnel fire smoke”
- Final report · Fire & disaster safety · 2023 Example hit 1 — theme only
- Talk · Fire & disaster safety · 2021 Example hit 2 — no document title
- Proposal · BIM policy & planning · 2020 Example hit 3 — fictional data
Project history · example
756 projects · 690 org names
- Find files — hybrid lexical/semantic search; pick by theme and year only.
- Project history — gather periods and orgs from the link table (still needs human check).
- Paper leads — skim nearby clusters on the meaning map.
- Performance lists — filter by type and theme into a list.
08
Who did what
People · Claude · Cursor · local models · code.
| Actor | Work |
|---|---|
| Person | Direction · spot checks · rule fixes · final judgment |
| Claude | Plan, judgment, result checks (Claude Code, Opus-class) · final review |
| Cursor | Scripting (Auto overnight) · Sonnet takes over when stuck |
| Local models | Bulk read/label/OCR/extract (gemma4:26b) · embeddings (bge-m3) |
| Code | Fingerprints · duplicates · MinHash · search index · RRF · UMAP · aggregates |
Hardware: 1× NVIDIA DGX Spark · evening 2026-11-09 → dawn 2026-11-11 · 3 overnight unattended runs