engblogs

A technical overview of Duolingo's approach to teaching speaking skills through integrated speaking exercises, mic-based practice, and real-time conversation tools.

Orchestrating travel planning with a three-agent architecture that uses Amazon Nova 2 Lite for reasoning and planning and Amazon Nova Act for browser automation to transform a fragile single-agent solution into a robust, explainable multi-agent system via simple A2A message-passing.

FAST provides a ready-to-deploy full-stack starter template for Amazon Bedrock AgentCore that accelerates agentic application development by wiring runtime, memory, tools, authentication, and observability into a production-ready stack with infrastructure-as-code via AWS CDK and a React frontend.

New Relic builds NOVA on AWS to transform internal productivity with generative AI, using Bedrock-backed RAG, Kendra, and MCP-driven Strands Agents to automate complex workflows across Confluence, GitHub, Slack, and Jira while maintaining security, sub-20-second latency, and measurable productivity gains.

Enterprise-scale LLM fine-tuning using Hugging Face and Amazon SageMaker AI combines distributed training, LoRA-based efficiency, and domain-specific data to deliver accurate, compliant, and cost-effective models.

MIT study reveals LLM ranking platforms are highly sensitive to a few votes, risking unreliable top models and prompting calls for more robust evaluation and richer feedback.

Unify SAP S/4HANA and Salesforce data on Databricks to deliver a governed, zero-copy, analytics-ready view of suppliers via Lakeflow and SAP BDC, enabling AI/BI dashboards and Genie-driven insights.

PART introduces a grid-free, self-supervised framework that learns the continuous relative arrangement of image parts to capture compositional structure, improving spatial understanding over grid-based methods and enabling universal pretraining across diverse data types.

Benchmark-driven, end-to-end guide to building production-ready Genie spaces that earn user trust through rigorous validation, explicit metadata, and domain-aware data practices.

VSSFlow unifies video-conditioned sound and speech generation through a flow-matching framework with a novel condition-aggregation mechanism, leveraging cross- and self-attention biases for V2S and VisualTTS in end-to-end joint learning to improve convergence and conditional generation.

Structured outputs on Amazon Bedrock enable constrained decoding to produce schema-compliant JSON responses, enabling zero-validation data pipelines, reliable agentic workflows, and production-ready AI applications.

An end-to-end guide to provisioning, configuring, and managing Amazon SageMaker HyperPod clusters with the HyperPod CLI and SDK, covering lifecycle automation, CRD-based workloads, and observability.