In the context of security, the use of AI agents is increasing, raising questions on how to maintain human oversight. Permissions alone do not ensure necessary judgement for each action. Repeated approvals for routine tasks can result in decreased attention and adherence to security. Amazon Quick's Smart Approvals addresses this challenge by reviewing eligible actions within existing controls, ...
In this post, we explore how Indegene, a strategic operating partner to the life sciences industry, built a large language model (LLM) powered Data Quality Agent within its NEXT Unified Data Platform on AWS. The agent generates context-aware data quality rules, validates them, and produces ready-to-run SQL and Python code under centralized, human-in-the-loop governance. Indegene's Data Quality ...
In this architecture, an AI agent application interacts with enterprise APIs using Amazon Bedrock AgentCore and Microsoft Entra ID. The goal is to maintain user identity throughout the request chain. AgentCore Gateway, a component of AgentCore, enables on-behalf-of (OBO) token exchange with Microsoft Entra ID. The AI agent authenticates with the gateway using a JSON Web Token (JWT) scoped to it...
Conversational AI, which includes virtual assistants, contact centers, telehealth, gaming, accessibility tools, and sales agents, is revolutionizing human-technology interaction. However, providing a genuinely natural voice experience is a significant challenge due to latency expectations. Users expect real-time response, and even slight delays can disrupt conversation flow and erode trust. Sta...
In today's digital age, building a professional reputation online is crucial for ambitious professionals. A well-crafted online voice can open doors and differentiate individuals in their fields. Norton Revamp, a generative AI-powered professional reputation management platform, addresses this need by providing real-time awareness of current trends in each user's professional domain. The platfo...
Advanced semiconductor manufacturing is becoming increasingly complex with shrinking error margins. Engineers need to accurately predict process behaviors, material interactions, and root causes of defects to optimize processes and minimize physical trials. AI is transforming semiconductor manufacturing through accelerated computing, physics-based AI, and digital twins. AI shortens simulation c...
Amazon Quick is a solution that allows organizations to connect and analyze both structured data from databases and spreadsheets, and unstructured data from documents and reports, within a unified experience. This is achieved through the use of a knowledge base for unstructured data stored in Amazon S3, and enriched datasets for structured data. Both types of data can be queried through a singl...
Energy experts gathered at the MIT Energy Initiative (MITEI) conference to discuss topics such as utilizing waste carbon at scale, assessing public trust in artificial intelligence and data centers, and building a quantum toolkit for critical applications. The conference attracted over 200 attendees and featured keynote speakers, panelists, seed grant innovators, and startup entrepreneurs. Spea...
At Circle K, a global convenience and fuel retailer, the high volume of alerts in their cloud-native infrastructure led to lengthy triage processes before remediation could begin. To streamline this process, the company built an automated triage system using Amazon Bedrock, a fully managed service that delivers foundation models through a single API. The AI-driven system autonomously investigat...
Sequence-level knowledge distillation (SeqKD) is a method used to fine-tune a shopping assistant on Amazon SageMaker AI. It uses a larger "teacher" model to generate high-quality responses, which a smaller "student" model is then trained on using supervised fine-tuning. SeqKD is cost-effective and portable as it requires no access to teacher logits or reinforcement-learning infrastructure. The ...
Amazon Bedrock offers Retrieval Augmented Generation (RAG) capabilities through its Managed Knowledge Base. You connect data sources, and Amazon Bedrock manages ingestion, storage, indexing, and retrieval. For web content, use the Amazon Bedrock Web Crawler connector. It crawls URLs or sitemaps you're authorized to access, respecting robots.txt. Custom preparation is necessary for specific sour...
This post discusses building a multi-agent engineering assistant using Amazon Bedrock AgentCore, Nova Multimodal Embeddings, and Strands Agents SDK. The assistant, designed to cut down manual cross-referencing in part substitution evaluation, is composed of three agents: Catalog, Evaluation, and Knowledge. The Catalog agent searches multimodal knowledge bases, the Evaluation agent scores candid...
A simple decision tree used for regression in machine learning will often overfit training data, resulting in near perfect accuracy on the training set but poor performance on new, unseen data. AdaBoost.R2, a version of adaptive boosting for regression, addresses this issue by creating an ensemble of many simple decision trees, each trained on a different subset of the source training data that...
Regina Barzilay and James Collins of MIT have been awarded the 2026 Pritzker Prize for their pioneering work on using AI to discover antibiotics. Their collaboration integrating deep learning, bioengineering, and experimental validation led to the discovery of halicin in 2020, the first artificial intelligence-discovered broad-spectrum antibiotic that effectively kills antibiotic-resistant bact...
Five MIT faculty members and two alumni have been named to the 2026 cohort of AI2050 Fellows, a philanthropic initiative aimed at solving hard problems in artificial intelligence. The fellows include Sendhil Mullainathan, a Peter de Florez Professor at MIT with a focus on human-centered AI, and Alex Damian, an assistant professor in the Department of Mathematics at MIT with a research focus on ...