Sample use case · Manufacturing
AI quoting copilot
From drawing to draft quote in minutes, built on ten years of past jobs.
New drawing
A-3121_mount-plate.pdf
no quote yet
K-0854-C_bracket.pdf
quoted CA$4,120 · margin 34%
P-0612_flange.pdf
quoted CA$3,860 · margin 29%
GR-1108_rail.pdf
quoted CA$5,340 · margin 31%
Drawing PDF
as it arrives
OCR + geometry parse
title block, materials, tolerances
Vector index of past jobs
text + geometric embeddings
Top-k similar jobs
with their real costs & margins
Copilot drafts the quote
retrieval-augmented LLM, every number traceable
Estimator signs off
adjusts, approves, sends
Retrieval does the heavy lifting: OCR and geometric embeddings put every past job in a vector index, and the copilot drafts from the closest precedents — so each line of the quote points back to a job the shop actually ran.
The problem
A precision machine shop quoted every job by hand: find similar past drawings, dig out old costs, rebuild the math in a spreadsheet. Three days per quote, and the knowledge lived in two senior estimators' heads.
What we build
- Indexed ten years of drawings and quotes with OCR — part numbers, materials, tolerances, final prices — into one searchable base.
- A similarity engine that surfaces the closest past jobs for any new drawing, with their real costs and margins beside them.
- A quoting copilot that drafts the estimate from those precedents; the estimator adjusts and signs off — every number traceable to a past job.
Proven before you commit
- Prototyped in three weeks on 300 historical drawings, blind-tested against quotes the shop had already sent.
- Ran four weeks in parallel with the manual process — estimators compared every AI draft against their own numbers before trusting it.
The outcome
3 days → 4 hrs
quote turnaround
92%
of drafts accepted with minor edits
2×
more quotes handled per estimator
A representative engagement — numbers illustrate the shape of the result, not a named client reference.