MSI Quote Studio
AI-enhanced quote estimator for custom manufacturing — complexity scoring, calibrated price recommendations, multi-line items, CAD upload with bounding-box extraction, branded PDF + email delivery. Inspired by Marking Systems Inc.
What it is
An AI-enhanced quote estimator for custom manufacturing — estimators describe a job in natural language, and the system returns a complexity score, a calibrated price range, and the reasoning behind it. Multi-line items, CAD upload with bounding-box extraction, branded PDF export, email delivery. Inspired by Marking Systems Inc.
The problem
Custom manufacturing quoting in the wild is two estimators in a Slack thread arguing about whether a Boeing nameplate run should cost $2,400 or $3,200. There's no shared reasoning, no historical anchoring, and the customer waits days for a number that ultimately came from someone's gut.
What I built
Complexity scoring
An LLM reads the job description (material, certs, tolerances, run size) and returns a 0-10 complexity score with a one-paragraph rationale. The score, prompt, and reasoning are logged for every prediction so estimators can audit them later.
Calibrated price band
Complexity score + run size feed into a calibrated regression that returns a suggested price range, not a single number. The range widens with score uncertainty — high-complexity jobs explicitly ask the estimator to confirm rather than auto-accept.
CAD upload + bounding box
Drag-and-drop a STEP / DXF and the system extracts a bounding box and surface area, which feed back into the quote (larger surface = more documentation overhead).
Branded PDF + email delivery
Once a quote is finalised, the system renders a branded PDF (cover sheet, line items, AI rationale appendix) and emails it to the customer with the estimator CC'd.
Engineering decisions
Why a price band, not a single number
Manufacturing pricing is genuinely uncertain. Returning a single number invites the customer to anchor on the bottom of the range; returning a band makes the uncertainty visible and shifts the conversation to 'why is this complex?' instead of 'why is this expensive?'
Why log every prediction
Estimators won't trust the AI without an audit trail. Logging prompt + score + rationale + final accepted price builds a calibration dataset I can use to retrain or fine-tune the regression once enough quotes have shipped.
Why bounding-box extraction over geometry parsing
A full STEP parser is a multi-week project. Bounding box gets you 80% of the pricing signal (volume, surface area, max dimension) for 5% of the work. A v2 can parse holes / threads / features.
What I'd do differently
Wire the calibration loop earlier. Right now logged predictions sit in a table waiting for me to write the retraining job; ideally the system would surface 'this prediction was off by X%' on the estimator dashboard within a week of launch. Also: build the CAD parser as a worker process from day one — synchronous bounding-box extraction blocks the request thread on large files.