Werk24 Significantly Accelerates Data Extraction from Technical Drawings: A Behind-the-Scenes Look
Introduction
At Werk24, we continuously strive to optimize the performance of our data extraction processes. Recently, we achieved a significant boost in speed for extracting information from technical drawings. In this blog post, we share how we managed to accelerate the process by focusing on three key areas:
1. Upgrading GPU Instances and Optimizing GPU Scheduling
One of the main bottlenecks in our previous setup was the processing capacity of our GPU instances. To address this, we upgraded to larger GPU instances, which allowed us to scale up the number of inferences that could be processed simultaneously. We also fine-tuned the GPU scheduling, which helped reduce the queuing time and ensured that the GPUs were working at their full potential.
As a result, our system can now handle more requests concurrently. However, while this upgrade significantly improved throughput, the new bottleneck has shifted to the bus between CPU memory and GPU memory. Needless to say, we’re already working on solutions to address this next challenge.
After resolving the GPU scheduling related issues in the first release (marked by the peaks in the middle), the second release allowed Werk24 to achieve a minimum of only 2.5 seconds for data extraction from complex techincal drawings!
2. Optimizing Code Paths for Targeted Feature Calculation
In any complex system, there is often redundancy in processing tasks. We identified and eliminated unnecessary feature calculations by optimizing the paths through our code.
This ensured that the system now only calculates the features that are strictly required for each specific request. This targeted approach reduces overhead, allowing for faster processing times without compromising the accuracy of our extractions.
3. Leveraging LLM for Code Reviews and Speed Optimization
To further improve performance, we introduced a Large Language Model (LLM)-based code review system. This AI-powered review process scans our codebase and automatically identifies areas where speed optimizations can be implemented. Beyond performance, it also highlights other potential improvements, ensuring that our code remains efficient and scalable as our systems grow.
Real-World Impact
The results speak for themselves. Our latest improvements have enabled us to process technical drawings at unprecedented speeds, with throughput increasing dramatically. For our users, this means faster quoting, more responsive procurement, and the ability to digitize spare parts and legacy documentation at scale. In some cases, AI-powered analysis can be up to 200 times faster than traditional manual methods.
What’s Next?
We’re not stopping here. As we continue to optimize our pipeline, we’re exploring new ways to further reduce latency and expand the range of features we can extract. Our goal is to make technical drawing data extraction as seamless and reliable as possible—so you can focus on what matters most: interpreting the data and making smarter business decisions. Stay tuned to our blog for future updates!
Ready to Accelerate Your Workflow?
If you’re still relying on manual processes or considering building your own extraction tools, let us show you what’s possible with Werk24. Our API is ready to handle your toughest drawings, and our team is always eager to extend our service to meet your unique needs. Get in touch to see how we can help you save time, reduce errors, and unlock the full value of your technical drawings.