Project

Structural Analysis Workbench (SAW)

Elbit Systems, UAV DivisionModiin, Israel

SAW is my desktop engineering system that takes a structural analysis project from the MSC Nastran input decks to a review-ready PowerPoint presentation and Word stress report: all the work after the solver went from days to a couple of hours, with one standard for the whole team. On top of it runs my agentic AI workflow in VS Code: AI agents plan, solve, post-process, review and assemble the whole project, while I make all engineering decisions.

I conceived, designed and built SAW alone, as the structural engineer, product owner, software architect and developer. I designed the AI skills that run my analyses through SAW, and I direct the team of AI agents that develops and tests it under my engineering rules. I use SAW on all my projects, and my stress team colleagues use it in their own projects.

days → hours
all the work after the solver, from days to a couple of hours
6
AI skills run a whole project, from the input decks to the report
5+
AI agents work in parallel on every run, 4 of them reviewers
57
automatic quality checks on every project
31 seconds
parallel post-processing of a real four-analysis project, 41 seconds more to build the documents
98k
lines of application code, covered by 102 test and benchmark scripts
SAW structure: methodology and data, the AI agent team that develops SAW, 5 analysis modules, the analysis skills and the SAW skill that orchestrates them, delivering presentations and reports
How SAW is built: from methodology and data to presentations and reports

What I did

  • Agentic workflow: I designed 6 AI skills that run a complete project: static, buckling, modal and random vibration, from the Nastran input decks to the final presentation and stress report, with almost no manual work
  • One working environment: I built the whole system in VS Code, and all the AI skills and agents run there today; every analysis and every correction is made from it, and the lessons of every project are written into the skill instructions, which the agents read on the next run
  • AI review board: I set up at least 5 AI agents to work in parallel on every run, 4 of them independent reviewers for model consistency, loads and spectra, solver health, and materials and allowables
  • Measured speed: I took a real four-analysis project through all solves in 3 min 19 s, parallel post-processing in 31 s and the build in 41 s, into a 37-slide presentation and a full stress report
  • Built-in quality control: I defined 57 automatic checks that flag poor solver accuracy, load imbalance, mass mismatch, low modal mass, resonance risk, high amplification, fatigue damage, negative margins and buckling below the applied load
  • One workbench: I brought the whole analysis process into one window with 5 modules (Post-Processor, Static, Fastener, Modal and Dynamic Analysis), one materials database and one method standard for every engineer and every project
  • One-click deliverables: I automated the company-format presentation and the Word stress report, with editable Office equations, numbered chapters, a margin of safety summary linked to every section and material appendices
  • No missed critical case: my 3D post-processor for MSC Nastran results checks every part in every load case, ranks the margins and picks the critical case of each part
  • Large-model performance: I made a 1.24-million-element model animate at about 60 frames per second instead of 1.7, with 62 optimizations, each proven to give the same numbers and the same picture
  • AI development team: I designed and direct 6 AI agents, from a CEO agent that plans and delegates the work to a QA engineer that independently reviews risky changes; they develop, test and document SAW under written engineering rules: about 98,000 lines of application code, covered by 102 test and benchmark scripts

Result

I turned days of manual work after the solver into a couple of hours, from the Nastran results to a checked stress report and presentation, with one standard for the whole team. My agentic AI workflow takes a whole project with almost no manual work, while I make all engineering decisions

Methods and tools

  • Agentic AI
  • VS Code
  • GitHub Copilot
  • Claude Opus
  • Python
  • PySide6
  • VTK
  • MSC Nastran
  • MSC Apex
  • Static
  • Fasteners
  • Modal
  • Random vibration
  • Fatigue
  • Buckling
  • Composites

Quality a reviewer can trace

  • Editable equations: every equation is a native, editable Office equation with the numbers substituted
  • Traceable allowables: every allowable carries its source, and the source tables go into the report appendices
  • Live margins: margins recalculate with every input, passing margins are green, failing margins are red
  • Consistent pictures: report pictures come out the same every time, an overall view, a part view and a zoom on the maximum
  • Established conventions: values agree with the established workflow, with the same result averaging as Altair HyperView
DADynamic Analysis: results exported from the 3D view, margins of safety checked, and a 26-slide presentation built in 15 seconds. Progress bars at 4× speed

Agentic AI: a whole project in one run

  1. 01

    Plan

    Every input is checked before any solve: analysis types, excitation directions, spectra, picture sizes and file path lengths

  2. 02

    Solve

    MSC Nastran runs one analysis at a time on a copy of the deck, and the original deck is never modified

  3. 03

    Fan out

    Analysis agents post-process static, modal, buckling and random vibration in parallel, while 4 reviewer agents read the model, the loads and spectra, the solver output and the materials

  4. 04

    Reconcile

    The orchestrating agent compares masses, frequencies and load cases across the analyses and decides every conflict

  5. 05

    Build

    The .saw project file, the full presentation and the full report, with a round-trip check that nothing was lost

Every solve is fingerprinted: when a deck or a spectrum changes, no report is built from stale results

Five modules, the whole analysis process

  • PPPost-Processor: a 3D viewer for MSC Nastran results that feeds every other module, generates the random vibration decks and runs the solver
  • SAStatic Analysis: margins of safety of metallic parts against MMPDS A- and B-basis allowables with thickness bands, and composite parts with the Tsai-Wu and maximum stress failure criteria
  • FAFastener Analysis: bolts, rivets and Hi-Lok pins in 7 failure modes, including preloaded tension with joint stiffness, bearing, pull-through and shear-out, from a library of 453 fastener-diameter records
  • MAModal Analysis: mode table, effective mass, animated mode shapes and an automatic resonance check against the vibration spectra
  • DADynamic Analysis: random vibration response and amplification in 3 axes, three-band fatigue with Miner's rule, and composite stress and strain margins

Engineering judgement stays with me

  • Human in the loop: the AI skills report numbers and flags; I decide every engineering question and sign off every result of my analyses
  • Same verified code: the automation drives SAW's own code exactly as an engineer would, so an automated report is as traceable as a manual one and reopens in SAW for review
  • Independent check: in the agent team, the one who builds is never the one who checks, and risky changes always get an independent QA review
  • Fast revisions: a model revision means regenerating the project, not rewriting it

All Elbit projects