engblogs

Explores Nano Banana 2, a Gemini 3.1 Flash Image model that fuses Nano Banana Pro's capabilities with Flash-speed generation, real-time web grounding, precise text rendering, and robust provenance across Google's ecosystem for high-quality visuals delivered at speed.

A technical overview of delivering Nasdaq eVestment data to the Databricks Lakehouse via Delta Sharing on Databricks Marketplace to enable real-time mandate discovery, AI-powered meeting prep, and automated workflow execution.

Deterministic automation via the Migration Intake and Processing Service (MIPS) replaces manual spreadsheets for Hyperforce migrations, enabling scalable, auditable decisioning across 95,000 organizations and 15,000+ requests with high-throughput auto-approval.

Explores how OpenAI Codex and Figma enable a seamless code-to-design workflow, bridging coding and design tooling.

An adaptive speculative decoding framework named Taming the Long Tail (TLT) trains a lightweight drafter on idle processors and uses an adaptive rollout engine to accelerate reinforcement-learning–driven LLM training, yielding 70% to 210% faster training with preserved accuracy.

Selective enhancement of math reasoning in LLMs by identifying pivotal tokens and circuits and updating only a sparse sub-network, achieving improved accuracy with minimal component changes.

Demonstrates efficient hosting of dozens of fine-tuned Mixture of Experts (MoE) models with multi-LoRA in vLLM on Amazon SageMaker AI and Amazon Bedrock, detailing kernel-level optimizations, tuned configurations, and latency gains (OTPS/TTFT) for GPT-OSS and related models.

PhysiOpt demonstrates how integrating physics-based simulations with generative AI can turn creative 3D designs into real, manufacturable personal items through rapid, constraint-aware optimization.

Deploy a production-grade, personalized event assistant at scale using Amazon Bedrock AgentCore and Knowledge Bases, leveraging memory, identity security, and serverless runtimes to deliver context-aware guidance for thousands of conference attendees.

A.R.I.S. uses a YOLOx-based deep-learning classifier to identify metals, plastics, and circuit boards in real-time, delivering a low-cost portable e-waste sorter that boosts material recovery with high precision and sortation purity.

A concise technical overview of approaches to detect, deter, and disrupt malicious AI usage in 2026.

An AI-driven multimodal framework that disentangles shared versus modality-specific information across cell measurements to reveal a holistic view of the cell state and its implications for disease.