I'm a structural engineer who builds software for the built environment. I train and evaluate computer-vision models on inspection photos, drawings and laser scans. I write openBIM tools that check and use IFC models, and I keep the engineering (Eurocodes, load paths, serviceability) in the loop.
My standard: results that survive review. That means field validation rather than benchmark-only scores, calculations reproducible by hand, tests that pin the numbers, and clear limits on what a tool may decide.
Currently
- Co-founder & CTO, AECAI Ltd: an AI-assisted structural-inspection platform (Dec 2025 onwards).
- R&D AI-Construction Specialist, AGECS: structural-drawing understanding and Scan-to-BIM (Apr 2026 onwards).
Before: 115+ structural and façade design packages as a graduate engineer at National Consulting Engineers (2023–24). MSc Digital Construction Analytics & BIM, Ulster University, Distinction (2025).
| Project | What it shows |
|---|---|
| concrete-defect-detection-shm | YOLO-seg vs U-Net vs FPN for concrete defects, then field validation on 44 real site images (recall 0.933, specificity reported honestly). Tested package, model card, leakage notes. MSc research |
| water-tank-crack-digital-twin | Calibrated crack width → severity class → Revit (Dynamo), Speckle, Power BI for a concrete water tank. Industry project with AECOM, 80% |
| Project | What it shows |
|---|---|
| ifc-model-auditor | BIM models tested like code: ISO 19650 naming, buildingSMART IDS, spatial and identity checks, duplicates, quantity take-off, schema-valid BCF 2.1, and a GitHub Action quality gate. Found 454 duplicated MEP fittings and a mis-mapped steel export in real Revit models |
| ifc-load-takedown | Column loads straight from IFC: Voronoi tributary areas, EN 1991-1-1 actions, EN 1990 6.10a/b with leading and accompanying actions, αn per category, equilibrium-checked, results written back into the IFC |
| ec2-crack-width | EN 1992-1-1 §7.3 crack widths and EN 1992-3 tightness limits, shown step by step. Cross-checked against fib's structuralcodes on 200 random cases. Compares measured crack widths with design limits |
| Project | What it shows |
|---|---|
| uk-construction-material-price-forecasting | 7 models on ONS and DBT material price indices, rolling-origin backtests against a no-change baseline, and P10/P50/P90 cost escalation with a coverage test. I found that my own coursework scores were in-sample and one was fitted on 11 of 132 months. A coverage test shows the 80% bands under-cover after the 2021–23 surge. MSc coursework, rebuilt |
| Project | What it shows |
|---|---|
| mohamedragab4554.github.io | My portfolio: Next.js + three.js with real-data 3D visuals (AI detections on structural plans, scan-to-IFC). Lighthouse 100 on desktop |
Professional R&D (not open source): structural-drawing understanding with an 8-class segmentation model (val mask mAP50 0.893), Scan-to-BIM from 250 M+ point laser scans to IFC, and a production inspection pipeline. These are employer and company work, shown as case studies on the portfolio only.
AI & data:
- PyTorch · segmentation-models-pytorch · Ultralytics YOLO · TensorFlow/Keras
- OpenCV · Open3D · NumPy · pandas · scikit-learn · statsmodels (time-series forecasting)
openBIM:
- IfcOpenShell · IDS / ifctester · BCF · Revit · Dynamo · Navisworks
- Speckle · Power BI · ISO 19650
Structural: Eurocodes (EN 1990, 1991, 1992) · AISC · ACI · ETABS · SAP2000 · SAFE · IDEA StatiCa · Tekla
Engineering software:
- Python · TypeScript / Next.js · pytest · GitHub Actions
- Docker · Supabase · serverless GPU inference
Metrics link to their source repositories. Academic work is labelled as such. Open-source tools are 2026 personal work built with AI pair-programming. Employer and client material is not published here.