Worldwide Express, a division of ShipStation Global, generated approximately $5 billion in revenue from over 130,000 customers and 75 LTL carriers. However, handling exceptions in their invoices resulted in $15 million in frozen free cash flow each month. These exceptions, categorized as line-item mismatches, pricing variances, and unexpected accessorial charges, required manual resolution thro...
Amazon Bedrock AgentCore is a platform to build and optimize coding agents at scale, allowing users to open and review web apps running in each session directly from their browser using Session Endpoints. Each session gets its own authenticated URL, and users can choose which ports to allow for incoming requests. With Session Endpoints, you can watch the webpage update in real-time as the agent...
Cost management for generative AI involves shared responsibility between application build teams and cloud spend accountable teams. Token consumption in generative AI scales with autonomy and context size rather than user interactions. Accurate cost tracking requires modeling rates and measuring actual consumption. The approach includes attributing every dollar, making requests observable, cont...
The ChatGPT desktop app, a desktop application of OpenAI's advanced models, is now accessible on Amazon Bedrock in AWS GovCloud (US) through two components: Work for larger tasks and Codex for technical work. The app is ideal for organizations with specific security requirements, routing model requests to Amazon Bedrock Mantle in us-gov-west-1 via AWS Identity and Access Management (IAM) and ex...
MIT researchers have developed a real-time analysis tool, called the LLM Election Observatory, to examine how AI chatbots respond to questions about the 2026 US midterm elections. The team, led by Professors Adam Berinsky, Chara Podimata, and Charles Stewart III, aims to understand how AI shapes the election discourse and influences voter understanding. The dashboard monitors how different AI s...
This post describes a method for migrating a mature PHP application to Java using an AI integrated development environment called Kiro. With Kiro, you can make PHP-to-Java migration a measurable and repeatable process. Kiro's approach centers on an engine with a shared methodology, mechanical discovery scanner, and an eight-phase workflow. The engine uses a parity harness to verify the accuracy...
A team has developed a novel approach to testing software called Just-in-Time Tests (JiTTests), which uses large language models to catch bugs on the fly. This method, which was previously shared as research, has now been successfully implemented at Meta, where it helps identify risks earlier and delivers reliable experiences at scale. JiTTests generate executable tests that demonstrate behavio...
MoEngage, a customer engagement platform used by over 1,350 consumer brands, faced a challenge when building AI agents that required access to internal APIs. These APIs, numbering around 400, already existed but needed to be safely accessible to agents without being exposed to the public internet, controlling identity, and without overwhelming context windows with hundreds of tool definitions. ...
A data scientist shares their experience of implementing kernel support vector regression (SVR) from scratch using Python, opting for stochastic gradient descent (SGD) instead of the commonly used SMO algorithm. They note the complexity of SMO and caution about incorrect information on SVR and SMO. The demo involves synthetic data, and the SGD-trained SVR model is shown to yield comparable resu...
AWS Batch now supports Amazon SageMaker Training jobs, enabling customers to manage and scale machine learning training workloads without building infrastructure. With the introduction of quota management in AWS Batch, administrators can divide and share GPU resources among teams or projects, maximizing utilization and prioritizing workloads. Quota shares include capacity limits and resource-sh...
NVIDIA's RTX Spark powers new thin and light laptops from major manufacturers, including ASUS, Dell, HP, Lenovo, Microsoft, and MSI. These systems are equipped with an NVIDIA Blackwell RTX GPU and Grace CPU in a single superchip, offering up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance. RTX Spark includes features like NVIDIA's CUDA platform, RTX graphics, and personal ...
The article discusses the challenge of scaling PyTorch training on a cluster, which is primarily caused by the complexity of building and maintaining the required containers. This involves ensuring that every container has the same CUDA runtime, PyTorch build, NCCL, EFA libraries, and AWS OFI NCCL plugin. To address this issue, AWS introduced the Ray Train Deep Learning Container (DLC), which i...
In the context of expanding quality assurance (QA) automation adoption within an organization, new challenges emerge as teams attempt to manage and organize an increasing number of tests for various applications. The author explores a solution called QA Studio, which facilitates defining tests in natural language and running them through AI-powered visual navigation. In this post, the focus shi...
Infrastructure has continually evolved in tandem with technological advancements throughout history. From the construction of dams, bridges, and canals during the 18th century to the development of railways, sewage systems, and megacities' infrastructure in the 20th century, each technological shift brought about new types and scales of infrastructure. Today, the definition of infrastructure ex...
In this post, we share the story of Property Finder, a digital real estate application, where 96.8% of their evaluation labels were found to be inherited from a single model's predictions rather than verified by human experts, leading to systemic bias and inaccurate comparisons of alternative models. Property Finder, which processes thousands of property images daily across the Middle East and ...