Today, the narrative around Artificial Intelligence (AI) is often driven by the bells and whistles of futuristic technologies. Think robodogs, AI chatbots, and the like. But AI isn’t new. It’s far-reaching yet often out of sight, transforming the everyday tools and platforms we interact with. To illustrate this, we’re sharing how Janea Systems contributed to three critical AI projects: PyTorch, Query Annotation Service (QAS), and Bing Spatial Data Services. They’re not quite as entertaining as a robot doing a backflip, but they’re definitely more useful.
Bing QAS: Optimizing Search Engine Accuracy
The Query Annotation Service (QAS) is integral to search engines, processing user inputs to identify the physical locations and entities users seek. As fundamental as it is, QAS had room for improvement, particularly in the accuracy of its geolocation results.
With our deep understanding of complex back-end systems, we were tasked to refine QAS and minimize these errors. Using various technologies, including C++, Python, and PowerShell scripting, we developed a script capable of batch-uploading query errors to an internal database. We created a refinement pipeline with smart features to notify stakeholders of any issues. The result? An improved QAS, powering multiple systems that rely on Bing geocoding.
Bing Spatial Data Services: Refining Geocoding
Bing Spatial Data Services play a pivotal role in geocoding – converting addresses into geographic coordinates. Janea Systems has been actively involved in maintaining and updating these services, contributing to three complex AI codebases: Bing Geocoding, the Bing Location & Understanding (BLU) Infrastructure Back-End, and the transition from the Extensible Application Platform (XAP) to Azure – the infrastructure for holding AI-based and general applications.
In the process, we’ve cleaned up the codebase, improved performance, modularized, and modernized the system to accelerate feature development and enable faster iteration on future proof of concepts (POCs). Our work has improved Bing’s geocoding services’ efficiency and helped uphold coding engineering guidelines, standards, and naming conventions.
PyTorch: Democratizing Machine Learning
PyTorch, developed by Facebook’s AI Research (FAIR) team, is an open-source machine learning library that plays a crucial role in fields like computer vision and natural language processing. Since its release in 2016, companies like OpenAI have adopted it extensively for various applications.
Despite its capabilities, PyTorch faced challenges related to compatibility and accessibility. Leveraging our system-level expertise, we ensure a stable version of PyTorch for Windows, enabling users to take advantage of new features introduced to C++17. We also contribute to the ongoing effort to port the library to run on Windows on devices with Arm-based processors.
We embarked on an extensive improvement process using a blend of technologies such as Python, C#, Arm64 architecture, Linux, the Gloo project, Continuous Integration (CI) tools, and .NET. Our team resolved Windows-specific CI test failures, assisted in transitioning PyTorch to C++17, implemented fixes in the Gloo project, and contributed to TorchSharp, extending the PyTorch API to the .NET ecosystem. These improvements have democratized access to PyTorch and reinforced its position as a go-to library for Machine Learning and Deep Learning, enabling the groundbreaking AI projects in the news today and in the future.
Before it even started making things, AI was a ubiquitous force, and these three examples provide a glimpse into the critical but often unseen roles that AI plays in our everyday interactions with digital technologies. As we continue to expand our AI practice with exciting projects and initiatives (such as our newest AI safety venture, BigFilter.ai), we remain committed to pushing the boundaries of AI, ML, and software engineering to create a better digital world.
Get in touch to find out more about our AI practice here.