Read 2D construction drawings and turn them into structured, decision-ready data — symbol and component detection, connectivity graphs, quantity takeoff, natural-language Q&A, and RFI drafts — across electrical, architectural, and P&ID disciplines, with results measured on real drawings against published state of the art.
| Capability | Benchmark | Result | State of the art |
|---|---|---|---|
| Electrical detection | DELP / SkeySpot (official split) | 0.847 ± 0.024 mAP@50 | SkeySpot 0.825 |
| Architectural detection | FloorPlanCAD (35 classes) | 0.820 ± 0.005 mAP@50 | — |
| Architectural (official split) | CubiCasa5K (9 classes) | 0.604 ± 0.001 mAP@50 | — |
| P&ID detection | PID2Graph OPEN100 (6 symbols) | 0.926 ± 0.008 mAP@50 | — |
| Connectivity (end-to-end) | PID2Graph OPEN100 edges | 0.752 edge AP | Relationformer 0.755 |
All numbers are 3-seed means ± std on held-out real test splits. Full methodology, splits, and
references: docs/BENCHMARKS.md.
A drawing flows through five layers, all reading and writing one schema — the Canonical Intermediate Representation (CIR):
- L0 · Ingest — PDF / DWG-DXF / IFC / image → gigapixel tiling → CIR
- L1 · Perception — symbol/component detection + connectivity-graph extraction
- L2 · Grounding — map detections to IFC classes + MasterFormat / UniFormat codes
- L3 · Engines — quantity takeoff (counts, areas, lengths)
- L4 · Agent — natural-language Q&A and RFI drafting over the CIR
Every extracted value carries a confidence score and a link back to its source entity + sheet.
Symbol & component detection — trained detectors on real drawings:
| Electrical | Architectural | P&ID |
|---|---|---|
 |
 |
 |
Connectivity graph (P&ID — symbols + junctions + pipe/signal edges) and quantity takeoff (from a real floor plan — real areas, counts, spec codes):
| Connectivity extraction | Quantity takeoff |
|---|---|
 |
 |
| Layer | Directory | Role |
|---|---|---|
| CIR schema | cir/ |
shared schema + (de)serialization |
| L0 ingestion + tiling | ingest/ |
PDF / DXF / IFC / image → CIR |
| L1 perception | perception/ |
detection + connectivity (docs/PERCEPTION.md) |
| L2 grounding | grounding/ |
IFC + MasterFormat / UniFormat codes |
| L3 engines | engines/ |
quantity takeoff |
| L4 agent | agent/ |
Q&A + RFI over the CIR |
| Evaluation | eval/ |
metrics, multi-seed CIs, cited SOTA baselines |
| Synthetic engine | synthetic/ |
IFC / parametric → rendered drawings |
| Backend API | backend/ |
FastAPI service + demo console over the CIR |
See docs/ARCHITECTURE.md and docs/CIR.md.
Requires Python ≥ 3.11 and uv.
uv sync --extra dev --extra eval --extra perception --extra backend
uv run pytest -q # test suite
# Quantity takeoff from a CIR document (counts + areas + lengths, spec-coded)
uv run python -m engines.takeoff --cir <file.cir> --format md
# Backend API + demo console (paste a CIR, run takeoff / Q&A / RFI)
uv run uvicorn backend.app:app --reload # http://127.0.0.1:8000The API also serves POST /takeoff-from-image (upload a drawing → detect → takeoff) and
POST /qa, POST /rfi. Point it at trained weights with
CDAI_DETECTOR_WEIGHTS='{"architectural": "<path>/best.pt"}'.
Benchmarks use publicly available datasets under their own licenses (attribution retained):
DELP (CC-BY), FloorPlanCAD (CC-BY-NC), CubiCasa5K (CC-BY-NC), PID2Graph (CC-BY-SA),
ResPlan (MIT). Data is versioned with DVC / kept local — never committed to git. See
datasets/registry.yaml.
Structural and civil disciplines (pending real detection data), combined-dataset training for cross-domain robustness, and expanded synthetic pretraining.
Free for noncommercial research and education under the PolyForm Noncommercial License 1.0.0. Commercial use requires a separate license — contact Alireza Shojaei (shojaei@vt.edu). Datasets and model weights remain under their upstream licenses.