Earth retaining systems — where the drawings meet reality
POSCO E&C · GeoAI · joint commercialization
Design documents and field instrumentation data are cross-verified, and a physics-informed AI ensemble (PINN, Bi-LSTM, XGBoost) predicts displacement. Review accuracy 99.2%; direct geotechnical analysis cost savings estimated at KRW 640 million per year. After 20 months of collaboration, the customer became a joint commercialization partner.
basis: POSCO E&C geotechnical review frozen gold set
Context — background
For a contractor, geotechnical and earth-retaining review meant collecting and cross-checking geotechnical investigation reports against field instrumentation data by hand. The reports run to hundreds of pages of PDF and the monitoring readings sit in a separate system, so cross-checking the two took a long time.
The existing workflow and its bottleneck
An average of 35 days per project to gather and reconcile the geotechnical review material — the depth of the review varied with each engineer's experience, and tracing the evidence behind a conclusion was difficult.
Data supplied
Geotechnical investigation report PDFs (borehole logs, material properties, field and laboratory test results) and field instrumentation data.
doAZ capabilities used
A deterministic extraction engine pulls soil layers, material properties and test values out of the report together with their coordinates on the source page, and a human-in-the-loop review workbench lets an engineer approve them value by value. The approved data builds a 3D stratigraphic model and sections, and a physics-informed AI ensemble (PINN, Bi-LSTM, XGBoost) predicts displacement.
What the AI does
Proposing extraction candidates, summarizing the site, and explaining risk factors — soft layers, groundwater, abrupt changes in strength — with the evidence behind each.
What rules, calculation and verification do
Ownership of the numbers. Validating extracted values, generating sections and models, and issuing verdicts are all performed by deterministic code. The LLM does not calculate.
Where a person reviews
In the review workbench, the source page is forced open beside every extracted value. The reviewer approves only after seeing the evidence, and the approval is recorded in an audit trail.
Measured outcome and its conditions
Geotechnical review 35 days → 30 minutes (95% faster), review accuracy 99.2%
Accuracy is measured against the POSCO E&C geotechnical review frozen gold set as the denominator — a structure in which cherry-picking is impossible.
Known limitations
The accuracy figure applies to the document types covered by the frozen gold set; we claim no generalization to every report format. On documents with poor scan quality the share handled by automatic extraction falls and the share requiring human review rises.
Customer approval scope
Customer name and figures approved for publication — a joint commercialization agreement was signed after 20 months of collaboration.
Other cases
We agree on the accuracy definition against a frozen gold set and then measure it — a structure in which cherry-picking is impossible.
Request a technical verification demo