The problem
Each price offer took two or three senior engineers three to five days. Roughly fifteen engineering-days for one bid. By the time it was ready the client had often signed with someone faster, and a single mistake on a 500-item bill of quantities could eat the whole margin.
What we built
Upload the BOQ and the system takes it from there: extract the line items, match each one against the client’s catalogue with GPT-4o and vector search over Qdrant, apply their pricing rules and margins, then stop. Pricing stays a human decision, so an engineer reviews before anything is generated.
Extraction, matching and document generation run as long jobs on Orleans grains, which is why the review screen never sits waiting on a 500-item file.
Inside the system
- Three days of work per offer became two hours. 85% faster.
- Five construction firms in Saudi Arabia, no pricing errors since deployment.
- Matching gets better with every offer the system completes.
- Full audit trail on every price calculation, for when a client disputes one.
- Separate catalogues and pricing rules per client.
Screens
