Methodology for PE-stampable AI
Active research threads on the methodology behind every claim in our product, tied to shipped code and aimed at peer-reviewed venues in structural engineering and construction informatics.
Large language models can draft calculations and propose designs. That is table stakes. The harder problem is building the governance, validation, and provenance layers that make the output something a PE can actually stamp. That is where we focus. Each thread below works on a piece of that problem, and every claim maps to code running in production today.
Research Threads
Six active threads spanning perception, design automation, provenance, human-in-the-loop constraint protocols, existing-building evaluation, and trust infrastructure.
Drawing-to-BIM perception that learns from your firm
Scanned PDFs, CAD exports, and marked-up plans all enter the same pipeline and come out as a structural BIM where every element tracks its origin. A frontier vision model produces the initial labeling; engineer corrections feed back into a detector tuned to your firm that improves with every reviewed drawing.
Available today in ImportModule and Design Studio.
Inverse design under validated code-check gates
Walk the AISC catalog lightest first and let validated calculators gate every section choice, so nothing passes that has not been checked. Underneath, an analysis ladder that scales fidelity to the decision at hand: closed form for concept screening, a differentiable surrogate for optimization, and full finite element verification before any commitment.
Available today in CalcModule and Design Studio.
A provenance architecture for AI-assisted PE work
Every reported quantity in a final calc package links back through a graph to the input artifact, code clause, calculator version, dataset version, the decision that selected it among alternatives, analysis tier, and the reviewer who signed it off. This is the technical foundation for PE defensibility when AI is part of the workflow.
Audit chain, PE review workflow, and Cmd-K dispatch audit are live in the platform.
Markup as constraints and auditable trade-offs
Instead of coaxing a generative system through chat, engineers draw on the plan in a vocabulary they already know. Those marks become hard or soft constraints for an automated layout solver, and competing schemes rank on a Pareto front where every metric (cost, embodied carbon, structural depth, schedule, code margin) carries a versioned, citable provenance record.
Markup-driven schemer and Pareto scoring are live in Design Studio.
End-to-end ASCE 41-23 Tier 2 evaluation
A full Tier 2 seismic evaluation of an existing reinforced concrete building, from E2K model import through component DCRs through stamped PDF, automated end to end. Deviations from published hand calculations are traceable to specific judgment inputs.
ASCE 41-23 calculator suite and packaged workflows are live in CalcModule.
Trust primitives for AI-assisted PE workflow
Engineer approval as a regulatory primitive, not a UX convenience. Scope statements and code edition metadata as audit fields on every calculation. Append-only history that cannot be rewritten. The architectural separation that lets PEs stamp output from an AI-assisted workflow without surrendering professional judgment.
The Assurance Stack is live across all modules.
Benchmarks We're Building
Validation corpora that back the empirical claims in each research thread.
- Anonymized real-firm structural drawing sets (5–10 sets) for perception and fine-tuning validation.
- AISC Steel Construction Manual worked-example replay set (10–15 examples from the F, E, and J series) for inverse-design validation.
- ASCE 41-23 Tier 2 benchmark buildings (3–5 buildings from PEER/ATC-76, NIST GCR 17-917-45, and FEMA P-2006).
- Cross-tier analysis benchmark (~30 regular steel buildings with T0 / T1 / T2 cross-validation).
Pilot firms willing to contribute anonymized drawing sets or benchmark buildings can reach out.
Scope Disclaimer
Research notes describe methods we use today and approaches we are exploring. Calculators in production carry scope statements and code-edition metadata; research material on this page does not constitute design guidance. Engineers remain responsible for verifying all results against the governing code editions for their project.