engblogs

Blueprint for reducing Salesforce Edge global rollback from hours to ten minutes for 1.5 trillion requests by deploying Kubernetes-native blue-green clusters, synchronized autoscaling, and automated traffic cutover with TCP connection draining to preserve four-nines availability.

A technical exploration of leveraging AI-powered personalized recommendations to optimize tax-deductible donations and donor engagement workflows.

VLSU presents a comprehensive framework to evaluate vision-language safety, revealing systematic joint image-text understanding failures, a fine-grained severity taxonomy across 17 patterns, and a large-scale benchmark of 8,187 samples to push robust multimodal safety research.

RA3 presents a scalable mid-training RL framework that learns compact action abstractions via a sequential variational objective to prune the action space, accelerate convergence, and boost post-training reasoning performance across code-generation and multimodal grounding tasks.

Exai Bio and Databricks demonstrate AI-powered, multimodal liquid biopsy for early cancer detection, detailing Exai-1 and Orion foundation models for cfRNA and oncRNA signals, unified on a Databricks Lakehouse with scalable compute and MLflow-based MLOps to enable reproducible, cross-cohort training across thousands of patient samples.

Databricks SQL 2025 delivers automatic, no-tuning performance gains across dashboards, pipelines, governance, and AI workloads, powered by serverless, Unity Catalog, Delta Sharing, and intelligent engine optimizations like Predictive Query Execution, while reducing storage costs with Zstandard by default.

Lambda joins the Open Compute Project Advisory Board to accelerate open, composable AI infrastructure that decouples power delivery and cooling from compute for scalable, repeatable deployments.

Explores whether LLMs can reveal their internal answer distributions, introduces the SelfReflect metric to gauge faithfulness between summaries and beliefs, and shows that explicit uncertainty requires sampling multiple outputs to be truly transparent.

An introductory overview of Prism, highlighting its core concept, envisioned use cases, and potential benefits.

Building a production-ready agentic AI solution on Amazon Bedrock using Pushpay’s GenAI evaluation framework, featuring a golden dataset, domain categories, and rapid iteration to reduce time-to-insight from minutes to seconds and reach 95% accuracy.

End-to-end full-stack Dart on Cloud Run, pairing a Flutter Web frontend with a Dart backend and shared data models in a monorepo workflow for scalable, serverless deployment.

ML infrastructure updates for Dataflow introduce new GPUs/TPUs, flexible provisioning, ML-aware streaming, and right-fitting heterogeneous compute to optimize batch and streaming workloads.