Industrial Subcontracting · Mechanical
From days to seconds to find a part that was already made
A 40-person mechanical subcontractor spent about a week quoting each part, mostly searching 100,000 PDF drawings for what had already been made. We built a prior-work search engine that finds the closest parts and the price actually paid, with zero generative AI.
<1s
prior-work search
18
title blocks, one engine
90%
fields read automatically
0
hallucination, deterministic
Before
- ✕Quoting a part took about a week, mostly spent finding what had already been made
- ✕Costing know-how lived in the memory of a few experienced buyers, written down nowhere
- ✕Around 100,000 PDF drawings, several parts per file, title blocks varying by client and year
- ✕The price paid lived in the supplier order, disconnected from the drawing it matched
After
- →Drop a drawing, get the closest historical parts in under a second (230 ms median over 10,000 parts)
- →18 different title block layouts read by a single engine, scanned drawings included, 90% of fields extracted
- →Shape signature invariant to rotation and symmetry: the same part turned 90 degrees is recognised as identical
- →119 of 136 parts linked to the price actually paid, with arithmetic checks on every order line
- →Per-field confidence score and mandatory human validation; every answer traces back to its drawing and order
- →Drawings are read then forgotten: only a thumbnail is stored, data fully exportable
Python 3.12FastAPIOpenCVPyMuPDFPostgreSQL (pgvector, HNSW)
Zero generative AI