Development of AI System for Construction Cost Estimation from Project Documentation
Construction estimate documentation comprises hundreds of volumes of technical documents: working drawings, specifications, work volume schedules. Manual cost estimation takes weeks; AI system reduces time by 5-10x.
Input Documentation
A construction project includes:
- Project Documentation: architectural section, structural, engineering systems
- Working Documentation: detailed drawings, details, equipment specifications
- Work Volume Schedules (WVS): tables with quantitative indicators
- Technical Conditions and SNiP/GOST requirements
AI Estimation System Architecture
[Load Documentation (PDF, DWG, Excel)]
→ [Section Classification]
→ [WVS Extraction (work volume schedules)]
→ [Equipment Specification Recognition]
→ [Apply Rates (FSNB, GESN, TER)]
→ [Direct Cost Calculation]
→ [Apply Coefficients (ISR, OZP, Overhead)]
→ [Summary Estimate]
→ [Human Verification]
Work Volume Extraction
class WorkItem(BaseModel):
section: str # project section
description: str # work name
unit: str # m², m³, units, linear m
quantity: float # volume
normative_code: str | None # GESN/FSNB code
confidence: float
def extract_work_volumes(document_text: str) -> list[WorkItem]:
# WVS tables - deterministic parsing
tables = extract_tables(document_text) # pdfplumber / Camelot
items = []
for table in tables:
if is_work_volume_table(table):
parsed = parse_work_volume_table(table)
items.extend(parsed)
# For unstructured descriptions - LLM
text_items = llm_extract_work_items(document_text)
return items + text_items
Normative Base Application
FSNB (Federal Estimate Standards Base) and regional TER collections contain rates for all types of construction work. System:
- Map work description → GESN/FSNB code via semantic search
- Obtain normative indicators (labor, equipment, materials)
- Multiply by current conversion indices to current price level (ISR)
Complexity: work descriptions in documentation don't always exactly match FSNB codes — normalization and expert mapping confirmation needed.
Equipment Recognition from Specifications
Equipment specifications (pumps, boilers, ventilation, electrical) — extraction with binding to supplier price lists or commercial quotes.
LLM extracts: name, model, technical characteristics, quantity. Next: request actual prices via supplier APIs or commercial search.
Verification and Audit
The system doesn't replace an engineer-estimator for complex projects — it speeds up their work. Each section of the estimate is accompanied by: data source, applied rates, calculation formulas. Full traceability for audit.
Accuracy and Limitations
AI estimate accuracy depends on documentation completeness. With full WVS set: deviation from manual estimate ±10-15% (acceptable for preliminary assessment). For tender documentation: ±5-8% with careful verification.
Timeline: MVP for typical residential building — 4-6 months; industrial object with FSNB and ERP integration — 8-12 months.