engblogs

MVICAD2 extends multi-view independent component analysis by modeling subject-specific temporal delays and dilations to reliably recover shared brain sources from MEG data, with identifiable likelihood and validation on Cam-CAN showing aging-related timing effects.

A comprehensive, large-scale evaluation of Group Relative Policy Optimization (GRPO) and Reinforcement Learning with Verifiable Rewards (RLVR) across non-English and multilingual settings, revealing crosslingual transfer gains, language-dependent trends, and implications for multilingual LLM reasoning.

A specialized semismooth Newton method reformulates kernel-based optimal transport as a nonsmooth fixed-point problem to enable scalable solutions with global convergence and local quadratic convergence, delivering substantial speedups over short-step interior-point methods.

MVICAD2 extends multi-view ICA by modeling subject-specific delays and dilations in MEG data, providing an identifiable likelihood and demonstrating superior cross-subject source estimation and aging-related brain dynamics on Cam-CAN.

A large-scale, multilingual study of GRPO and RLVR across non-English settings reveals crosslingual gains, native-language reasoning efficiency, and the need for broad evaluation to detect language-specific regressions.

Case study on deploying secure, multi-tenant AI agents on AWS using Amazon Bedrock AgentCore, featuring per-tenant runtimes, strict tenant isolation, and integrated cost governance for Axonius.

Axonius demonstrates a secure, silo-based multi-tenant AI agent deployment on AWS using Bedrock AgentCore, detailing per-tenant isolation, cost governance, and built-in observability.

AIDA on Amazon Bedrock Knowledge Bases uses implicit and explicit filtering and metadata-enriched chunking in a RAG workflow to dramatically improve contract search accuracy and deliver grounded, context-aware answers.

Jumio's streaming-first, real-time feature store on AWS demonstrates a scalable architecture using Kinesis, Apache Flink, and SageMaker Feature Store to achieve sub-100ms latency for fraud-detection ML and streamlined feature management.

A multi-agent document classification system on Amazon Bedrock that fuses Claude Haiku 4.5-driven textual reasoning with Titan Multimodal Embeddings for visual pattern recognition, orchestrated by the Strands Agents SDK for cross-validated, auditable results.

Customize and brand the Amazon Quick embedded chat in your app by aligning visuals and tone through container CSS, frame options, and agent persona with SDK content options and sendPrompt integration for a seamless, native-like experience.

AgentCore payments is GA, enabling autonomous, pay-per-use transactions for AI agents with wallet integration, protocol-agnostic payment orchestration, guardrails, and production observability at scale.