Meeting with Tayyab¶
Original OneNote page: Survey and research - Meeting with Tayyab
Route: meeting/person note
Categories: people_meetings
Source¶
- Word export:
C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\TextCAD_Wiki\inbox\onenoteexport\01_PROJECTS\Gauss Compute\Survey and research - Meeting with Tayyab.docx - MHT export:
C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\TextCAD_Wiki\inbox\onenoteexport\01_PROJECTS\Gauss Compute\Survey and research - Meeting with Tayyab.mht
Extracted Notes¶
These notes are extracted from the Word export so the source is visible in the wiki without moving or deleting the original file.
- Meeting with Tayyab
- For part level 1000-2000 drawings
- Restricted to part level drawings
- They have around 2200 drawings
- For better results, get more data
- Natural of drawings matters
- Multi agent systems - drawing + step - develop a multi agent system for tool selection
- First started with 400 drawings
- Notes
- Do they template match? Raw vision matching
- Template matching can also be done depending upon the problem
- 9 labels
- Extending bounding boxes
- If giving a small bounding box its easier, than instead of the whole drawing
- Assembly level drawings - more than one yolo model
- One model detects the key views
- Some views are merging with each other
- At this stage, still exploring assembly level drawings with merging views
- Local SME drawings are very complex
- Automotive or ship building
- If we get some views, 10 views, based on prompt enginering give to the vision language model, cropped bounding box
- Generic knowledge of vlm to tell
- Asking vlm to check if this info is complete of incomplete
- If the info is being cropped
- Don’t need to fine tune a vlm (gpt 4o)
- First model will get 10 views
- Then input it to vlm, then ask the vlm if the bounding boxes are cropped, unique view (a bit confusing)
- No need to use the opv model (because views are already along x-y axes)
- For bounding boxed, inclindation yolo v model
- For paper analysis
- If the size of checkpoing is higher, thenn inference speed lower
- Yolo v11m,
- Industry more interested in x version, has more accuracy
- For vision language model
- Florence, open source (general purpose model)
- Donut (ocr free model)
- Donut is better, only for information extraction (from the drawing for us)
- Check latest model at huggingface
- Since working in a government institute, they approach the university for this support
- DATA
- ASME textBooks
- FOR CAD Models
- Method
- GNN based (developed model to extract key features, machining features)
- For complex like shaft, not suited
- Python OCC, algorithm to develop geomterical information of
- Not good, but for higher level
- Convert complex parts to snapshots
- Ask VLM what features do you see
- Most of the papers, detection of features
- Industrial problem to extract dimensions as well, raidus or orientation of the hole
- Use mathematical based scripts to extract geometric features
- Big misconception
- AI cant do wont 100% job
- Llm and vlm skip some information
- He has developed a multi-agent system
- Applied human in the loop
- Review some annotations and correct it
- Manufacturing is precision driven industry, not completely rely on AI
- Hyundai
- Machining as a case study
- More mature system, 100-200 years
- Working in A-start and NTU
- May be we can collaborate on NTU side
- Win-win situation could be a collaboration
- If not on university level, we can also do consultation
- Professor: Moon
Next Curation Action¶
Promote this into a person page, link it to the relevant company/account, and create/update the EspoCRM contact if this relationship is still active.
Sources¶
C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\TextCAD_Wiki\inbox\onenoteexport\01_PROJECTS\Gauss Compute\Survey and research - Meeting with Tayyab.docxC:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\TextCAD_Wiki\inbox\onenoteexport\01_PROJECTS\Gauss Compute\Survey and research - Meeting with Tayyab.mht