engblogs

Analyzes the P-Completeness of evaluating nested Boolean query DAGs on inverted indices and introduces ComputePN, a sparsity-aware evaluator with a Positive-Negative dual representation that bounds time to O(|Q| · |U_active|) and avoids exponential blowups and the universal scan.

A concise multi-dimensional analysis of human-like behaviors in LLMs, examining model behaviors, user factors, and system prompts, backed by large-scale evaluation across multiple models to illuminate controllability and guide responsible LLM design and evaluation.

Decision framework for when to refit, respecify, or hold production forecasting models in the face of drift and shocks, illustrated by COVID-era disruptions at Airbnb.

A practical framework for deciding whether to refit, respecify, or hold a production forecasting model, using COVID-era lessons from Airbnb to balance data drift, structural change, and risk.

KnowledgeForge demonstrates an AI-driven pipeline that mines resolved ITSM tickets to automatically generate and curate high-quality knowledge-base articles, with deduplication, quality scoring, and a closed loop that feeds back into generation, all orchestrated on AWS Bedrock, S3 Vectors, and Step Functions.

Fanatics Betting and Gaming’s scalable multi-agent AI system on AWS orchestrates specialized agents via EKS and Bedrock, with Guardrails, MCP tool integration, and a custom RAG pipeline to deliver fast, compliant, state-specific answers and seamless human escalation during live sports events.

Explores asynchronous invocation patterns for Amazon Bedrock AgentCore agents in serverless pipelines, illustrating how task-token callbacks, direct Step Functions integration, and durable functions reduce idle Lambda costs while coordinating long-running AI reasoning.

A scalable, serverless blueprint that automates mortgage document intake end-to-end—classifying, extracting, and validating data with the AWS GAIC IDP Accelerator and Quick Automate to orchestrate downstream workflows, delivering faster time-to-close and lower costs.

Explains runtime domain and published-date filtering for Web Search on Amazon Bedrock AgentCore, detailing the admin-plus-runtime policy model, per-call API usage, and regional availability.

Turn trusted business context into production-ready, domain-specific AI agents from a single prompt by grounding them in Unity Catalog data and documents, then scope, benchmark, and expand with evolving knowledge and workflows.

Asana uses OpenAI Codex and frontier models to replace an outdated Enzyme testing system, delivering in about two weeks what was expected to take five years, at a fraction of the cost through AI agents and automated codebase migrations.

NVIDIA scales expertise by turning ChatGPT Work into automated, reusable workflows that surface actionable AI signals, cut manual tasks, and accelerate prototypes and global planning across GTM and solutions architecture.