engblogs

Cross-model disagreement paired with a total uncertainty metric combines epistemic and aleatoric uncertainty to more reliably detect overconfident, unreliable LLM predictions using a diverse ensemble and fewer queries.

Generative AI-enhanced wireless vision uses mmWave reflections to reconstruct hidden objects and rooms, boosting precision, privacy, and robotic manipulation.

Explores how Agentforce converts LLM outputs into structured UI to power rich, interactive agent experiences across 4M sessions, balancing automation, context preservation, and safe, adaptable interfaces.

Migration-focused guide detailing how to replace Airflow with Databricks Lakeflow Jobs, translating XComs, sensors, backfills, branching, and dynamic task mapping to data-first, lakehouse orchestration.

A technical overview of Facebook Reels' Friend Bubbles system, detailing ML-based viewer–friend closeness, content relevance ranking, a continuous feedback loop, and performance-conscious metadata integration to enhance social discovery and engagement.

Goldilocks RL proposes a teacher-driven data sampling strategy to tailor task difficulty for reasoning in language models under sparse rewards, enabling more sample-efficient learning with a Goldilocks-inspired curriculum and GRPO.

FabCon 2026 showcases Azure Databricks' Lakeflow, Lakebase, and Genie innovations—unifying data ingestion, operational data, and AI-assisted analytics on an open lakehouse foundation with deep Microsoft 365 integrations.

Part 1 shows how to build a modular Multi-Agent System (Dev Signal) that discovers Reddit questions, grounds findings with Google Cloud docs via MCP and the Google ADK, and generates blogs and visuals, deployed to Google Cloud Run with a long-term memory layer for future expansion.

Practical guide to scaling read workloads with Cloud SQL read pools and autoscaling, enabling a single read endpoint, automatic node adjustments, and cost-efficient high availability.

GA of Memorystore for Valkey 9.0 delivers massive throughput and lower latency with SIMD optimizations, pipeline memory prefetching, new HEXPIRE commands, geospatial querying, and cluster-enabled multi-database support for scalable, managed caching.

Orchestrates a distributed team of specialized AI agents—researcher, judge, and orchestrator—into scalable microservices deployed on Cloud Run, enabling seamless frontend integration via the ADK and A2A protocol.

Nova Forge SDK unifies end-to-end LLM customization for enterprises, enabling developers to prepare data, train models, and deploy Nova variants across Amazon Bedrock and Amazon SageMaker AI with guided workflows and scalable automation.