Written Twice, Read Once: SeoulTech's Pipeline Checks Calculation Against Drawing, No BIM Needed
SeoulTech's document-AI pipeline matches calculation and drawing by member code, flagging 253 discrepancies on one project, no BIM model required.
Take a column. Call it C12. It appears in the structural calculation report with its section, its main bars and its stirrups. It appears again in the member schedule on a structural drawing, and the bars there should match. Two documents, one building, made by different people at different moments. Over a project, they drift apart.
A team at Seoul National University of Science and Technology (SeoulTech) has tackled exactly this. Under Associate Professor Jaewook Jeong of the Department of Safety Engineering, lead author Minsang Gu and colleagues built a pipeline that reads both documents, pairs every member, and lists where the two copies disagree. The paper, “Detection of human errors in construction structural design documents using automated extraction algorithm”, appears in Engineering Applications of Artificial Intelligence (2026). As Tech Xplore reported on 3 November, Jeong sums it up plainly: “Our AI-assisted framework automatically extracts, organizes and compares information.”
A design is a promise written twice
A structural design exists twice. The calculation is the proof that the member carries its load. The drawing is the instruction the site builds from. Coordination is the work of keeping those two copies equal, and every Bauleitung team knows the moment they stop agreeing. SeoulTech’s contribution is to make that comparison mechanical.
The machine, in five steps
- Find the regions. On the calculation documents, a deep-learning layout model, DocLayNet-YOLO, locates the blocks that hold member data.
- Read the text. Surya-OCR extracts the values.
- Find the tables. Drawing schedules have rows and columns of uneven width, so OCR is combined with edge detection, morphological operations and a Hough transform. That transform goes back to Paul Hough’s 1962 patent, generalised by Richard Duda and Peter Hart in 1972: a quiet tool generations of engineers have relied on without knowing the names behind it.
- Match on a key. Member code plus direction joins the two documents. The system also flags duplicates and sorts the mismatches by type.
- Ship it. A web application merges the extracted data into one file.
This is not a chatbot and not a language model. It is classical document AI followed by rules: narrow, auditable, and working on the documents offices already exchange.
←TODAY: In 2026, a SeoulTech pipeline reconciles calculation and drawing by member code and direction. →3012: Every load-bearing element carries one identity from first sketch to demolition record. Fulcrum: The check only becomes a lookup when someone gives the member a name both documents print.
What the list showed
Gu et al. report that on a residential project in South Korea, the pipeline flagged 253 discrepancies: 224 missing entries, 15 main-reinforcement mismatches, 3 stirrup mismatches, 8 dimensional discrepancies and 3 duplicates. Most of the list is about presence, a member that exists in one document and not in the other. The breakdown maps where coordination leaks.
The trade-off sits in the accuracy figures. According to the SeoulTech team, extraction reached 95.56 % on calculation documents and 90.47 % on drawings. The drawing side is the harder read, so a person still reviews every flagged line before anyone acts. The pipeline shortens the search. It does not sign the plans.
Jeong is explicit about the scope: “Importantly, the system performs these checks without relying on building information modeling (BIM).” He names the target too: “It could be particularly valuable for small- and medium-sized construction projects that cannot implement BIM.” The tool does not compete with BIM. It reaches the projects BIM has not yet reached.
The Swiss desk
Many small and mid-size Swiss projects, a Mehrfamilienhaus extension or a school annex, still move between Bauingenieur and architect as PDF sets. That is exactly where this check matters. openBIM answers the problem with IFC and machine-checkable IDS requirements; underneath both answers sits the same primitive, a shared, printed key on every member. PAZ has covered that thread before: Sebastian Lieb’s BSS thesis, FOREST & HUMAN, built a “place of truth in Grasshopper” linking Archicad, RFEM and Excel so revisions propagate without re-entry. A model that carries member IDs makes reconciliation trivial. A plan set without IDs turns it into archaeology.
Atelier: An office living with AI checkers will find that the machine is only as good as the keys the office prints. If C12 is labelled “C12” in the calculation and “Stütze 12” on the plan, no pipeline can pair them. Monday move: add a Member ID property to every structural element in your Archicad or Rhino template and show it in the plan label.
Hack: Join your own calculation and schedule exports on member code plus direction, the same key SeoulTech uses. Export both lists (an Archicad schedule, a Grasshopper data tree, or the engineer’s spreadsheet) into SQLite 3.39+ or Postgres and run:
SELECT COALESCE(c.code, d.code) AS code, COALESCE(c.dir, d.dir) AS dir, c.main_bar AS calc_bar, d.main_bar AS dwg_bar
FROM calc_members c FULL OUTER JOIN dwg_members d
ON c.code = d.code AND c.dir = d.dir
WHERE c.code IS NULL OR d.code IS NULL
OR c.main_bar <> d.main_bar;A NULL on one side is a missing entry; the remaining rows are reinforcement mismatches. A second pass with GROUP BY code, dir HAVING COUNT(*) > 1 finds the duplicates.
From where I sit, half a century on, the offices that coped best were not the ones with the most software. They were the ones whose members had names. Give yours one before the next plan issue.
PAZ Kaffi · multidisciplinary editorial, led by PAZ Academy