The AI reads the PDF print into proposals an estimator confirms
VenturusEQ sends the PDF drawing sheet to Anthropic's Claude and gets back the title block, material, dimensions, holes with threads, tolerances with datums, finishes, heat treat, welds, operations and notes, each with a confidence. Every value is a proposal your estimator confirms, and the confirmation is written to a ledger. The AI never prices anything.
What comes back from the print
Every field below comes back as a proposal with a confidence.
Identity
- Part number
- Revision
- Title
- Sheet manifest: sheets declared and sheets read
- Variant tables
Material and treatment
- Material
- Mass
- Thickness
- Finish
- Heat treat
- Special processes
Features on the print
- Dimensions
- Holes: size, depth, thread, quantity, bolt circle
- Welds
Quality
- Tolerances with their datums
- Surface finishes
Process
- Operations called out
- Notes
Confidence
- A confidence on every value
- A value the read is unsure of is flagged, never replaced with a made-up number
Oversized sheets are read in overlapping crops
An oversized sheet goes as overlapping crops of one document, so a note in the corner is not lost. The read reports sheets declared against sheets read, the part scope, and any anomaly. Where the read is unsure of a value, the line says so rather than inventing one.
Up to 25 files per read
Up to 25 files go in per request and 100 can wait per company. A read is retried on failure and can be canceled. Thread call-outs come from the print alone; a tapped hole is never guessed from the solid.
From a read value to a confirmed fact
The software reads and proposes, a person confirms, and only confirmed values are priced.
The read lands as proposals
Each value arrives on the line with its source and a confidence. Nothing is written to the quote as a fact, and nothing is priced.
The estimator confirms, corrects or ignores
A confirmed value is ledgered with its source, the method that produced it, its confidence, who confirmed it and when. A corrected value is ledgered the same way with the person as its source.
Part numbers are offered, never overwritten
When the title block disagrees with the line, a pill offers Relabel or Keep and names the evidence it read. The line keeps its number until a person chooses.
Only then can the value be priced
A formula that reaches for a value nobody has confirmed fails rather than pricing quietly. The confirm step is what separates a read value from a priced one.
What is sent to the AI and what never leaves the product
Export-control boundary
Switched on per deployment. A file classed as controlled is refused before it leaves for the read, and text derived from a controlled file asks the same per-person switch as the bytes.
Geometry never leaves the container
The CAD kernel measures STEP, IGES, DXF, DWG and STL inside the product's container. No model goes to the AI. Only the PDF sheet does, and only for the drawing read.
No training on your data
The product's security and data ownership statement says it plainly: customer data is not used to train models. Anthropic is named as the subprocessor for the read.
The ledger
Every confirmation, relabel and controlled open is written down with who and when, so the record shows what a person accepted, not what the AI proposed.
Security and data ownership →What the AI never does
- It never prices a part. Prices come from your rates, your formulas and your confirmed inputs.
- It never picks a process template. Templates are activated by your administrators and chosen by your estimators.
- It never writes or changes a rate. The product ships no rates and the AI cannot add one.
- It never selects a method. Method proposals from geometry are offered, never auto-selected.
Common questions
Which AI reads my drawings, and what exactly gets sent?
Anthropic's Claude. The PDF drawing sheet itself is sent as one document, with an oversized sheet cut into overlapping crops so nothing at the edge is lost. Back comes a structured set of proposals: title block, material, dimensions, holes with threads, tolerances with datums, finishes, heat treat, welds, operations and notes, each with a confidence. No model and no rate is sent.
Can it read a drawing that has three sheets?
Yes. The read reports sheets declared against sheets read, so a sheet that did not come through is noticed rather than missed, and the PDF viewer shows a rail of sheets and remembers where you were on each one. An oversized sheet is read in overlapping crops so a note in the corner is not lost.
Is my drawing used to train an AI?
No. Customer data is not used for training. The product's own security and data ownership statement says so and names Anthropic as the subprocessor for the drawing read.
What happens when the AI reads a value wrong?
Nothing, until a person acts. A read value is a proposal on the line with its confidence. The estimator confirms it, corrects it or ignores it, and the confirmation is written to a ledger with the source, the method, the confidence, who confirmed it and when. A wrong proposal nobody confirmed can never reach a price.
We do ITAR work. Will those drawings be sent to the AI?
Not where the export-control boundary is switched on for your deployment. A file classed as controlled is refused before it leaves, and text derived from a controlled file asks the same per-person switch as the bytes. Geometry is measured inside the container and never goes to the AI at all.
See it on one of your own RFQs
Send us a drawing and a model you have already quoted and we will walk through it with you in the product, in your rates. If you would rather look first, the interactive demo is open, with no form in front of it.