AI quoting from STEP files and PDF drawings for CNC job shops, on your own server

What AI quoting from drawings does

AI quoting from drawings reads a customer's request for quotation, the 3D model as a STEP file and the 2D drawing as a PDF, and prepares the costing for the estimator: material and blank size, the features that drive machining time, critical tolerances and surfaces, and the most similar parts your shop has made, with what they cost. The estimator checks the proposal, corrects it and decides the price. For a contract machining, sheet-metal or mould shop, the GPU work is the reading of drawings, and one NVIDIA RTX PRO 6000 Blackwell card does it for the request volume of a whole shop.

This article is for the owner or managing director of a job shop and for the person who prepares quotes. It explains each step in plain terms, says which parts need a graphics processor and which run on any computer, and puts the hardware in one section with one table.

The pressure is familiar. ki-syndikat.de, a German catalogue of AI use cases for small and medium-sized companies, puts manual costing of complex CNC turned and milled parts at 1 to 4 hours per request. A request that never becomes an order takes the same estimating time as one that does. The same page is careful about the gains: it calls the time saving clear and makes better accuracy depend on calibration and human review.

Confidential RFQ drawings and NDAs: why job shops keep them in-house

Customer drawings are the customer's property and often come with a non-disclosure agreement. The ki-syndikat page on CNC quote costing states that a self-built solution on the shop's own infrastructure, where the data never leaves the server, is often the only solution accepted by customers in defence, medical technology or the automotive OEM field. A shop that uploads such drawings to an external quoting service may breach its contract even if the service is well secured.

Hosted quoting services differ in where they process data. The same catalogue lists services hosted in the EU with a data processing agreement and services that process data in the USA, and recommends a legal review where customer contracts restrict transfers outside the EU. Read the confidentiality clauses of your largest customers before choosing; your legal counsel confirms what they allow. Our article on trade secrets, NDAs and AI tools covers the business side, and our guide to the CLOUD Act and EU data residency explains why the location of a provider's company matters as well as the location of its data centre.

With your own server, drawings stay in the network where your CAD and ERP data already live, and the question of third-party processing does not arise.

From STEP file and PDF drawing to the quote in four steps

Steps as described on the ki-syndikat.de page on CNC quote costing (read 25 October 2026); "runs on" is our assessment from the documentation of Open CASCADE Technology 8.0.1, XGBoost 3.4.2 and the model card of Qwen3.8-27B, all read 25 October 2026.

The 3D model. A STEP file is the exchange format of 3D CAD (ISO 10303). Feature recognition software reads it and finds holes, threads, undercuts, grooves, fits and contours, which feed estimates of machining time, set-up time, programming effort, material and tool wear, as the ki-syndikat page lists. This is geometry processing on the processor. The open-source Open CASCADE Technology library, for example, reads STEP files according to AP214 and AP203 and "some parts of AP242" in its version 8.0.1 documentation, including assembly structures, and handles colours, layers and GD&T through its extended data tools.

The 2D drawing. The PDF drawing carries what the 3D model often lacks: the material and its standard, the general tolerance class, individual tolerances and fits, surface roughness, heat treatment, coating and the quantity. The ki-syndikat page warns that 2D drawings are processed with lower accuracy and that GD&T annotations in a PDF may be missed if they are absent from the product and manufacturing information in the STEP file. Reading this information is the job of a vision-language model, an AI model that looks at the page as an image and writes out what it sees as structured text. This is the step that needs a GPU.

Similar past parts. Your ERP holds every part you have quoted and made, with the planned and the actual times. Finding the ten most similar parts by material, size, feature counts and tolerance class is a database query over numbers the first two steps produced. It runs on the processor and gives the estimator a direct reference: what a similar part cost you in the end.

The cost model. A machine-learning model can learn the relation between part features and actual times from your post-calculation. For such a self-built model, the ki-syndikat page names at least 200 to 300 comparable orders as the data volume needed and, in another section, 300 or more, so plan on the higher figure. It names gradient-boosting methods such as XGBoost and LightGBM. A data set of a few hundred orders is small for these methods, and they run on an ordinary processor; XGBoost can also train on a GPU, which its documentation enables with device set to cuda.

We supply workstations and servers with RTX PRO 6000 cards, and our Private AI/ML service sets up the drawing reader and its ERP connection; send us a few typical requests and the number of quotes per week, and our team replies with a proposal.

Reading 2D drawings with a vision-language model

Qwen3.8-27B is an open model that its card describes as "a native vision-language model that understands images and videos", published under Apache 2.0. Its FP8 version has a checkpoint of 30.9 GB. The model card reports document results such as 91.1 on OmniDocBench 1.5 and names no tests on technical drawings, so the accuracy on your drawings is something to measure.

How the model sees a drawing depends on resolution. Its preprocessor makes one image token per block of 32 × 32 pixels and accepts up to 16,777,216 pixels per image.

Patch size 16 and merge size 2 (one token per 32 × 32 pixels) and the pixel limit from the preprocessor_config.json of Qwen/Qwen3.8-27B on Hugging Face, read 25 October 2026; pixels and tokens are our arithmetic.

An A1 drawing at 200 dpi has about 31 million pixels, so the processor scales it down to the limit, which is about 147 dpi by our arithmetic. Small tolerance figures in a crowded view can become hard to read at that resolution. A practical pipeline therefore sends the title block, the notes field and each view as separate crops at full resolution, and asks the model for one structured answer per crop.

Three settings matter. The model runs in thinking mode by default, and the card allows it to be switched off per request with enable_thinking set to False, which suits extraction. The serving engine vLLM can force the answer into a JSON schema, so the result arrives as fields: material, standard, tolerance class, each tolerance with its feature, roughness, treatment and quantity. And every extracted value is shown to the estimator next to the drawing, because a misread tolerance makes the quote wrong.

What a test looks like. Take 50 drawings from recent requests, including handwritten changes, poor scans and drawings from different customers' CAD systems. Have an estimator mark the correct values once, run the model and count the errors per field. The result tells you which fields the model can prefill and which the estimator still types. The ki-syndikat page adds a practical rule: a short internal checklist, six points are enough, that every estimator goes through before releasing a quote prepared with AI.

The hardware a job shop needs

The drawing reader is the only part that needs a GPU, and one card is enough for a shop's request volume. By our users-per-card rule, 90 per cent of the memory the driver reports less 3 GiB per card, one 96 GB RTX PRO 6000 holds 83.0 GiB of weights and cache, as our guide on how many users one RTX PRO 6000 serves derives. The FP8 weights take 28.75 GiB and leave 54.3 GiB for the cache.

Our guide to vision-language models for document AI puts this model's cache at 64 KiB per token in 16-bit plus a fixed 144 MiB per request, which matches its configuration of 16 full-attention layers with four key-value heads of 256 dimensions. With the cache stored in FP8, as in all our sizing, that becomes 32 KiB per token. An A3 drawing of about 7,500 image tokens with about 800 tokens of answer, our assumption, then takes about 0.39 GiB, so one card holds about 137 drawings in flight by our estimate. A job shop that receives a few dozen requests a day uses a fraction of that.

For one drawing on its own, the speed limit is the memory bandwidth: the card reads about 28.4 GB of weights per generated token, so the RTX PRO 6000 Server Edition at 1,597 GB/s writes at most about 56 tokens per second and the Max-Q at 1,792 GB/s about 63, by our estimate an upper bound. An answer of 800 tokens then takes about 14 seconds plus the time to read the image.

Capacities by our users-per-card rule; power by our estimate (cards, processor, plus 15 per cent for the rest of the system) as in our Max-Q workstation guides and the entry configuration on our AI servers page, all read 25 October 2026.

One RTX PRO 6000 Max-Q in a workstation. The NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition is a 96 GB card with a 300 W blower cooler. In a workstation it runs the drawing reader and the cost model on an ordinary 230 V office circuit, at about 0.75 kW at the wall by our estimate. It suits a shop that wants to start with quoting alone.

Two Max-Q cards in a workstation. The second card carries a language model for the estimating team, for example to search earlier quotes and customer correspondence, at about 1.1 kW. Our guide to buying a 2× RTX PRO 6000 Max-Q workstation lists the parts.

A two-card server. The entry server on our AI servers page has two RTX PRO 6000 Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives, at about 1.8 kW by our estimate. It belongs in a server room and serves several departments: quoting on one card, the company assistant, order entry from emails and PDFs and document search on the other. For continuous operation in a rack we recommend the Server Edition; the Max-Q belongs in a workstation or in a rack chassis whose maker lists it.

Notes for the IT team

1. Where the files are. Requests arrive by email or through a customer portal; keep the STEP and PDF files in the document management or ERP system that already holds them, and let the pipeline read them from there.
2. Geometry. STEP reading and feature recognition run on the processor. Open CASCADE Technology 8.0.1 reads AP214, AP203 and parts of AP242; check which AP your customers' CAD systems export and whether their PMI carries the tolerances.
3. Serving. Run the vision-language model with vLLM behind an LLM gateway with keys and a query log; set enable_thinking to False in chat_template_kwargs for extraction and pass a JSON schema for the fields. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere (vLLM 0.32.0 security guide), so keep the engine on an internal network.
4. Software support. The RTX PRO Blackwell cards are compute capability 12.0; the driver, CUDA and engine releases must support them, as our guide to RTX PRO Blackwell software support explains.
5. Cost model. Export planned and actual times, material, batch size and the extracted features of past orders from the ERP; retrain when the post-calculation shows systematic deviations. The ki-syndikat page suggests a quarterly review.
6. Paper archives. Older drawings on paper or as raster scans can be brought into the reference set; our article on digitising paper drawing archives with AI covers that work.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build workstations with one or two RTX PRO 6000 Max-Q and servers with RTX PRO 6000 Server Edition cards to order, assembled and tested before delivery, delivered within the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, delivered with our engineering partner Vixen.UNO, sets up the drawing reader, the search over past orders and the connection to your ERP; describe your parts, customers and CAD formats, and we propose a pilot with your own drawings.

FAQ