It reads the CAD drawings that arrive — DXF, vector PDF or a scan — and reviews them for errors and missing information, compares revisions, reconciles the drawing set against the BOM and the purchase order, and fills the quality paperwork that goes with the job. Every finding carries a confidence score, a location on the sheet and a human approval gate.
Built and delivered for a tier-2 precision machining and sheet metal supplier working to AS9100 / ISO 9001, reviewing drawings from roughly 40 OEM customers every week.
Overview
This is a long page on purpose — it documents a delivered system rather than a concept. Jump straight to the part you are evaluating.
Context
New and revised part drawings arrived every week from around 40 OEM customers. Most of an engineer's day went on checking prints and retyping the same paperwork.
Where the time went
Title block, units, dimensions, tolerances, GD&T, notes and every referenced spec, read line by line on every sheet that came in.
Two PDFs overlaid on top of each other and the difference spotted by looking. Slow on a good day, and unreliable on a dense sheet.
Inspection plans, control plans, routers, PPAP forms and certificates of conformance, typed again from the drawing for every job.
Usually after the material had already been ordered, which is the expensive moment to discover a disagreement.
The single largest source of scrap and customer nonconformances was a job running to the wrong revision.
What the quality manager asked for
The constraint that shaped the design
The shop did not trust a black box. Deterministic checks had to be explainable line by line, and anything the model interpreted had to carry a confidence score and a location on the sheet. That one requirement is why the arithmetic lives in Python and the language model is never asked to do it.
What it does
Each capability is a separate tool-calling node in the agent graph. Engineers use them one at a time, or as a single intake pass over a new drawing package.
Title block completeness, unit consistency, missing or duplicated dimensions, tolerance stacks that do not close, GD&T frames with missing or undefined datums, thread callouts against hole size, and notes referencing specs that are not on file. Each finding has a confidence score and a marked-up PDF.
A Rev C to Rev D change report: which dimensions moved, which notes were added or deleted, which tolerances changed, whether material or finish changed, whether the change affects fit, and whether the ECO number on the sheet matches PLM.
Reads the BOM, the drawing set and the purchase order together. Flags quantity mismatches on sub-assemblies, hardware in the BOM that is not shown on the print, and material or finish callouts that differ between the PO and the drawing.
Fills your Word and Excel templates — inspection plan, control plan, router, PPAP, CoC — from the drawing and the BOM. Restructures or bulk-edits templates from a plain-language request, with tracked changes and an approval step.
Shop floor and office staff ask questions in plain language and get an answer with the source sheet and view cited. Available in the web UI and through an Amazon Lex chat channel.
Anything that can be computed is computed in plain Python: unit maths, tolerance stacks, datum existence, thread versus hole tables. The model handles reading, interpretation, ambiguity and writing. It is never asked to do arithmetic.
Product tour
A real shop sends a bit of everything: a vector PDF with a STEP model, a native DXF, a two-sheet assembly, a 3D cross-check and a scan from 1998. Part numbers and customer details below are sample data.
The queue
Every uploaded drawing with its job, part number, revision, input format, which pass it is on, how many findings are open and when it last changed. Mixed input is normal here: vector PDF, DXF, DWG with a STEP model, a TIFF scan still in extraction.
Turned part · vector PDF with STEP linked
A vector PDF with the STEP model linked. 23 dimensions, one feature control frame and five notes were extracted, then checked in 36 seconds.
Sheet metal · native DXF
Parsed natively with ezdxf: 19 dimensions, one bend table and five notes, checked in 22 seconds. The flat pattern is recomputed by the rule engine rather than trusted.
Assembly · balloons and parts list
Two sheets with the BOM and the purchase order linked. The parts list is parsed as a table, then cross-checked against BOM rev 3 and PO 88213.
3D · drawing against STEP model
The STEP AP214 model matched against the drawing in 58 seconds, feature by feature, at 0.05 mm size and 0.5 mm position tolerance. Nine features match, three do not.
A 1998 weld neck flange arriving as a TIFF goes through the raster side of the pipeline first. PaddleOCR plus a YOLOv8 layout model trained on a few thousand annotated sheets locate the title block, the notes and the callouts, and each detected region is shown with its own confidence before a single rule runs. Geometry checks do not run on a scan — there is no geometry to check — but the title block, notes and callouts are read and checked like any other sheet. Budget about 20 seconds per scanned sheet, and 300 dpi straight scans give the best results.
Who it is for
Built for manufacturing, industrial machinery and equipment — anywhere a drawing package is the contract between a customer and your shop floor.
Job shops and tier-1/tier-2 suppliers taking prints from many OEMs, each with its own drafting conventions and title block.
Assemblies with deep parts lists and long revision histories, where a balloon without a parts list entry becomes a missing part on the line.
Bend tables, flat patterns and hole-to-bend rules that have to agree with each other before the laser runs.
Where every finding needs to be traceable, reviewable and reversible, and the paperwork has to be generated from the drawing rather than retyped.
Decades of scanned sheets that nobody wants to re-draw, but that still have to be read and checked when a repeat order lands.
If your largest single cause of scrap is a job that ran to the wrong revision, that is exactly the problem this was built to remove.
Check catalogue
The deterministic checks run in Python, so they return the same result every time and every finding names the rule it came from.
This is the general drafting check list. Checks specific to your drawing standard and your own company check list are added during the project — that is the part that gets adapted, and it is where most of the value is for a shop with its own conventions.
Revision compare
Not a pixel diff. Every change is read as a change to a dimension, a note, a tolerance or a material, and then classified by the department it lands on.
| # | Item | Rev C | Rev D | Change | Impact |
|---|---|---|---|---|---|
| 1 | Corner holes, 4X | Ø6.35 +0.05/-0 | Ø6.60 +0.05/-0 | +0.25 diameter | FIT M6 clearance changed, mating bracket 4471-021 should be checked |
| 2 | Overall length | 120.0 ±0.2 | 120.0 ±0.1 | Tolerance halved | INSPECTION CMM instead of caliper, inspection plan characteristic 3 |
| 3 | Slot length | 22.0 | 24.0 | +2.0 | FIT Adjusts stud travel |
| 4 | Note 3 | TAP 4X M6x1.0-6H | — | Removed | PROCESS Router step 30 (tap) no longer needed |
| 5 | Note 6 material | 6061-T6 | 6061-T651 | Temper | PURCHASING PO 88213 still lists T6 |
| 6 | Material note | — | AL 6061-T6 PLATE, 6.0 THK | Added | COSMETIC Duplicates the title block |
Inspection plan IP-4471-020 needs two characteristics updated and router step 30 (tap) should be removed. The change report exports to xlsx, opens as a side-by-side highlighted PDF, and can be sent straight to the quality inbox.
Drawing vs BOM vs PO
The drawing set, the BOM export and the purchase order are read together as one package. Four conflicts on a three-drawing job is a normal morning.
| Field | Drawing | BOM | PO | Status |
|---|---|---|---|---|
| Qty, PEM nut CLS-M4-1 | 4X (sheet 2) | 6 | 6 | QTY MISMATCH |
| Material, 4471-020 | AL 6061-T651 | AL 6061-T651 | AL 6061-T6 | PO OUTDATED |
| Finish, 4471-020 | Anodize Type II clear | — | Chem film | FINISH CONFLICT |
| Washer NAS1149F0432P | not shown | 8 | 8 | MISSING ON PRINT |
| Revision, 4471-020 | D | D | D | MATCH |
| Qty, 4471-021 sub-assy | 2X | 2 | 2 | MATCH |
Sources: 4471-000 Rev B sheets 1-2, 4471-020 Rev D, 4471-021 Rev A · BOM export from Epicor · PO 88213 PDF
STEP model cross-check
STEP AP203 and AP214 models are parsed with pythonOCC and matched feature by feature against the hole callouts and diameters on the drawing.
Holes, bores, counterbores, pockets and overall envelope are extracted from the solid and paired with the callouts on the sheet. Matching runs at 0.05 mm on size and 0.5 mm on position.
A feature whose size or depth differs, a feature that the drawing calls out but the model does not contain, and a feature in the model that the drawing never dimensions.
The model is shown rotatable with each feature coloured by result: matching features neutral, mismatches red, features present but not called out in amber.
The most common real finding is a model that was never re-exported after the drawing changed, or a hole edited on the print but not in CAD.
Quality documentation
Nothing is written to the shared drive without a human accepting it. That was a hard requirement from the quality manager, and it is what made the rollout stick.
Characteristics pulled from the sheet with nominal, tolerance, method and frequency, each carrying the sheet and view it came from. Rows that rest on an unresolved finding are flagged rather than filled.
Generated as docx from your own CoC template, with the finish specification and revision taken from the drawing rather than retyped.
One section per open finding, with a callout illustration generated for each so the operator sees the feature being described.
Filled from the drawing and the BOM into the same template set you use today, as docx and xlsx.
A quality manager can ask, in one sentence, to update every template that references QS-114 Rev B to Rev D and add a Cert on file checkbox to the CoC under the finish section. The agent finds every occurrence across the template set, shows a preview of how many it will touch, and writes Word files with tracked changes and Excel files with a change-log sheet. Approve all, review one by one, or reject.
Q&A agent
A second agent graph in which the model calls tools: it searches the drawing package and the spec index, reads findings and reports, and cites the source of every answer.
Shop floor
What is the surface finish on the bore?
Ra 0.8 µm (32 µin) on the Ø25.00 H7 bore. The callout is on 4471-021 Rev A, sheet 1, Section A-A. Note 2 on that sheet adds: no tool marks visible after anodize.
4471-021 Rev A · Sheet 1 · Section A-A · Note 2
Purchasing
Which drawings in this job call out anodize?
Two of three. 4471-020 Rev D (Type II clear, note 4) and 4471-021 Rev A (Type II clear, title block). 4471-000 Rev B has no finish callout because it is an assembly drawing.
4471-020 · Note 4 — 4471-021 · Title block
Shop floor access
A drawing tool that only lives on an engineer's screen misses the moment that matters — the machinist at the setup, the inspector at the CMM, the storeman unpacking the wrong temper. Two channels bring them in without asking anyone to learn a new system.
The Q&A agent is exposed through an Amazon Lex chat channel, so a question can be asked from a shop terminal or a handset instead of a logged-in browser. The answer is the same one the engineer gets, with the sheet, view and note cited — Lex carries the conversation, the agent does the reading and the citing.
Typing into a tablet with gloves on, next to a running machine, is the reason observations end up on paper and then nowhere. An inspector records a few seconds of speech against a finding, AssemblyAI transcribes it, and the text is attached to the log where that finding lives — searchable, and part of the audit trail rather than a note in someone's pocket.
Questions that get asked away from the desk
Machinist at setup
What is the finish on the bore?
An answer at the terminal instead of a walk to the office — or a guess that turns into rework.
Setter starting a job
Is this job running the latest revision?
A job running to a superseded revision was the single largest cause of scrap on the original project. The machine is the last place to catch it.
Inspector at the CMM
Which characteristics here are 100% inspection?
The inspection plan was generated from this drawing, so the answer comes from the same source rather than a printout that may be a revision behind.
Purchasing
Which drawings in this job call out anodize?
Finish and material conflicts between the drawing and the PO surface before the order goes out, not when the parts come back wrong.
Planner
Does this part still need the tap operation?
When a revision removes a note, the router step it justified is usually still there. The change report already knows; now the planner can ask it directly.
Goods-in
What material and temper should have arrived?
6061-T6 and 6061-T651 look identical on a label. Catching the difference at the dock is far cheaper than at final inspection.
What a voice note actually gets used for
Why it is worth wiring up
The hard part of a rollout is never the engineering team — it is everyone downstream of them. A question answered in a channel people already have open costs no training and no new habit.
Questions, answers and voice notes land in the same audit log as the findings. Under AS9100 or ISO 9001 that is the difference between a conversation and a record.
A transcribed note explaining why a finding was wrong goes into the index of past decisions, so the same false positive is less likely on the next drawing from that customer.
Which chat channel the agent is delivered into, how long voice audio is retained and which languages are transcribed are decided during deployment against your IT and quality policies. They are not fixed by the product.
Inputs and outputs
Every input format is read by the method that suits it, so a native DXF is never downgraded to an image and a scan is never assumed to carry geometry.
Input · how it is read
2D drawings, vector PDF
Text and geometry extracted directly; a layout model locates the title block, dimension strings, GD&T frames and notes
2D drawings, scanned or raster PDF / TIFF
PaddleOCR plus a custom YOLOv8 layout model trained on a few thousand annotated sheets
Native DWG / DXF
Parsed with ezdxf at entity level: dimensions, blocks, attributes, layers, MTEXT. Model space and paper-space layouts both read
STEP models (AP203 / AP214)
pythonOCC geometry cross-checks: does the dimensioned hole exist at that size and position
BOM (xlsx, csv, ERP export)
Structured parse, part numbers matched against the ERP
Purchase orders (PDF)
Line items, material, finish, quantity, revision
Supplier manuals, internal specs, ASME Y14.5 excerpts
Chunked and embedded in pgvector for retrieval at check time
Existing Word / Excel templates
python-docx and openpyxl, with tracked changes on edit
Output · format
Findings report with severity, confidence and location
JSON + marked-up PDF overlay
Revision change report
Table + side-by-side highlighted PDF
Cross-document discrepancy report
Table, exportable to xlsx
Filled inspection plan, control plan, router, PPAP, CoC
docx / xlsx from your own templates
Template edits and restructures
docx / xlsx with tracked changes, pending approval
Answers to plain-language questions
Text with sheet / view citation
Audit log
Every finding, every accept / reject, every file touched
| File you upload | What it reads | Depth |
|---|---|---|
| DXF (AutoCAD 2013 or later) | The most complete review: title block, notes, dimensions with their geometry, GD&T, threads, holes, parts list, balloons, sheet-metal tables. | Full |
| PDF plotted from CAD | Title block, notes, GD&T, callouts, threads and circles from the text layer and vector geometry. Overridden and duplicated dimensions need dimension objects, which a plot does not carry. | High |
| Scans (PNG, JPG, TIFF, image-only PDF) | Title block, notes and callouts by OCR and the layout model. 300 dpi and a straight scan give the best results. Geometry checks do not run. | Text only |
| STEP (AP203 / AP214) | Holes and bores compared with the hole callouts and diameters on the drawing. | Geometry |
| BOM and PO (xlsx or csv) | A header row with a part-number column and a quantity or description column. | Structured |
| DWG | Not read directly — it is a closed binary format. In AutoCAD use Save As → DXF (or DXFOUT) and upload the DXF instead. | Export first |
Under the hood
Deterministic where it can be, model-driven where it has to be, and a human approval gate before anything leaves the system.
Every drawing runs through a graph — ingest, extract, rule check, retrieval enrichment, finalize — on background workers. The pipeline view shows each stage and its time for every drawing, so a slow or failed step is visible rather than mysterious.
The checks themselves: title block, units, datums and GD&T references, threads against hole sizes, dimensions, parts list, revisions and specs. These results do not depend on the language model, so they are repeatable, and every finding shows the rule it came from.
A vector index holds the spec library, customer rules, ASME Y14.5 excerpts, the document templates and past review decisions. In the enrichment step each finding is linked to the clause or spec it rests on.
A second graph in which the model calls tools: it searches the drawing package and the index, reads findings and reports, and cites the source of each answer.
One drawing through the system
Upload
PDF, DXF or STEP in, job linked from the ERP
Extract
Layout model plus OCR, or native entity parse
Rule check
Deterministic checks in Python, model interprets the rest
Enrich
Retrieve specs, customer rules and past findings
Finalize
Findings and markup to the viewer
Review
Engineer accepts, documents generated, audit logged
Stack
Integrates with
Accuracy and limits
About 8% false positives after threshold tuning with engineer feedback. Here is the honest account of how that number was reached and what it costs to get there.
What to expect in the first weeks
Title-block wording differs per customer and is set up once. DXF export settings decide what can be checked. Scans work best at 300 dpi. The spec register and released revisions have to stay current, ideally read straight from your ERP or PLM.
Expect to want checks from your own drafting standard, your own document templates, adjusted severities and batch intake of drawings. All of that is normal and planned for.
False positives on unusual drawing styles in the first weeks, reduced by tuning against your engineers' accept and reject decisions, plus OCR errors on poor scans. This is the part that needs your engineers in the loop.
Per-customer setup
Every OEM prints its title block differently. Rather than guess, the agent shows you what it read and asks you to name it — then re-checks every drawing from that customer.
Spec numbers, revisions and whether each is current or superseded, as CSV or Excel. It goes into the index immediately, and drawings are then checked against it — a note calling a superseded revision becomes a finding.
The released-revision list so the agent knows which revision each part should be running to. In a project this comes directly from your ERP or PLM instead of a file.
The agent lists what it read but did not recognise — for example the label SURFACE TREATMENT and the block attribute TREATMENT, both holding HARD ANODIZE, with the field Finish suggested. Accept the suggestions in one click.
Every drawing from that customer is re-checked. Setup findings disappear, real drafting findings stay, and any decisions you already made are kept.
A new customer's nine-drawing package is uploaded before any setup exists. The agent returns findings on every sheet, including two that are only about the missing setup: a title block field it cannot place and a part it cannot find in PLM. After two minutes of setup and a re-check, the finish is read on all nine drawings, the correct spacer drops to zero findings, the genuine drafting problems remain, and one new and correct finding appears — a note calling a spec revision that the customer's own register marks superseded.
How we deliver it
The ingestion pipeline, agent graph, rule engine, review UI, markup overlay, template editing, approval gate and audit log carry over. Your standard, your title blocks, your specs, your templates and your ERP are what get adapted.
Collect sample drawings, your template set and the exact check list you want enforced. This is also where we establish which standard you work to — ASME Y14.5, ISO GPS or a company standard — and which formats your drawings actually arrive in.
Entity-level parsing for native CAD, OCR plus layout detection for scanned sheets, and a title-block profile per customer so the fields land where they should.
The checks that must be exact are written as code. Model integration handles interpretation and writing, never the arithmetic.
A readable change report with impact classes, a markup overlay, and BOM and PO reconciliation wired to your ERP.
Your document templates connected, edits produced with tracked changes, and nothing written back until a human approves.
A feedback loop that tunes false positives against your engineers' real accept and reject decisions, on your real drawings.
Estimate
3 to 4 weeks
Typical end-to-end delivery for a first production agent, including tuning. The scope that moves this number is the number of integrations and how many exceptions your check list contains.
Measured after rollout
~60%
less engineering time
Time spent on incoming drawing review, measured by the quality team after the agent was in regular use.
~1/3
fewer nonconformances
Drawing-related nonconformances fell by about a third once revision and cross-document checks ran at intake.
~8%
false positive rate
On findings, after threshold tuning against real engineer accept and reject decisions.
How it was rolled out
Discovery: sample drawings, templates, check list
Ingestion and rule engine validated on real sheets
Pilot with two engineers, findings side by side with manual review
Tuning from accept / reject feedback
Full team, plus shop floor Q&A
Try it before you commit
About 45 minutes gets you through a reference package and a customer the system has never seen. After that you can load your own drawings.
A drive-module package is already loaded: an assembly with a STEP model, BOM and PO, the same drawings as PDF plots, a sheet-metal bracket and a scanned sheet. You see the expected findings on each, accept and reject them, export the markup PDF and the findings CSV, and look at the extraction and pipeline detail behind a check.
Upload a nine-drawing package from a fictional customer with its own title-block wording, part numbers and spec numbering. Review what it finds before setup, then import the spec register and released revisions, teach it the wording and re-check. This is the part that shows what onboarding a new OEM actually costs you.
Your formats, your title blocks, your part numbers. Expect some low-severity findings about specs and revisions the test instance cannot know about, and expect to add your own title-block wording once.
Good to know
ASME Y14.5, ISO GPS or your own company standard — that choice is part of the adaptation in a project. The general drafting checks (title block, units, datum existence, thread versus hole, duplicated dimensions, parts list, sheet metal) run regardless. Excerpts of the standard you work to are loaded into the retrieval index so each finding can cite the clause it rests on.
No, and we would rather say so plainly than pretend. DWG is a closed binary format. In AutoCAD use Save As → DXF (or DXFOUT) and upload the DXF — the review is then the most complete of any format, because dimensions come through with their geometry attached.
They go through OCR and a layout model that locates the title block, notes and callouts, each with a confidence score you can inspect before any rule runs. Text-based checks all work. Geometry checks do not, because a scan carries no geometry. Budget about 20 seconds per sheet and scan at 300 dpi.
Never without a human accepting it. The approval gate was a hard requirement from the quality manager on the original project and it is built into every write path: documents are produced with tracked changes, shown for review, and only then written back. Every finding, decision and file touched is in the audit log.
About 8% false positives after tuning with engineer feedback, on the delivered project. Deterministic checks — unit maths, tolerance stacks, datum existence, thread against hole tables — are exact because they are code, not a model. Findings that require interpretation carry a confidence score and a location on the sheet so an engineer can triage them in seconds. Expect a higher rate in the first weeks, on unusual drawing styles, until thresholds are tuned on your sheets.
No. Deployment into your own infrastructure is the standard option: HTTPS, your login or SSO, backups and monitoring, with the source code handed over. Where drawings go is your decision, not a constraint of the product.
Three to four weeks for a production deployment including ingestion, your check list, your templates, ERP or PLM integration and tuning on your real sheets. You can see the generic build working on your own drawings within an hour on the test instance, before committing to anything.
About two minutes in the browser. Import their spec register and released revisions, confirm the title-block wording the agent could not place, and re-check. From then on their drawings are read automatically.
That is what the chat channel and the voice notes are for. The Q&A agent is exposed through an Amazon Lex chat channel, so a machinist or an inspector can ask from a shop terminal or a handset and get the same cited answer an engineer gets in the browser. Inspectors record voice notes against a finding, which are transcribed and attached to the log. Which channel it is delivered into is decided during deployment.
Yes, and this is usually where the most value sits for a shop with its own conventions. Rules like minimum hole-to-bend distance for a given thickness and radius, required fit classes on keyways, or mandatory chamfer notation are written into the rule engine during the project so they are checked exactly and identically every time.
No. It does the mechanical pass — the reading, the cross-referencing and the arithmetic — and hands a ranked, evidenced list to a person who decides. Every finding is accepted or rejected by a human, and those decisions are what tune it. The measured result on the delivered project was about 60% less time spent on review, not a role removed.
Keep going
Manufacturers usually automate one expensive bottleneck first, then add the next once it has paid for itself.
Consistent quotes in minutes, not evenings
Quoting is where margin quietly leaks. We encode your price list, volume tiers, discount limits and approval thresholds once, and the agent applies them the same way every time — producing a branded quote document from an enquiry while the customer is still interested.
More infoSupplier feeds to live listings, no retyping
For shops that add products faster than a person can describe them. The agent ingests supplier feeds in whatever shape they arrive, writes descriptions that read like a human wrote them, translates them into every market you sell in, and pushes complete, SEO-ready listings into your store.
More infoAutomated email handling with your tone of voice
Your shared inbox stops being a queue. The agent reads each incoming message, classifies it, pulls the customer's history from your CRM and drafts a reply that matches how your company writes. Routine confirmations go out on their own; anything sensitive waits for one click from you.
More infoTell us which standard you work to and which formats your drawings arrive in, and we will run them through the agent and walk you through what it found — including the findings it got wrong. That conversation is free, and you keep the results.
Contact usOr just call us
Pick a slot that suits you — 30 minutes, online, no preparation needed.
