GlycemicGPT brings agentic AI to diabetes care

TLDR: Open-source diabetes agent GlycemicGPT lands, AWS ships a blueprint for real-time voice agents, and Amazon Lex gets a safer NLU upgrade.

GlycemicGPT launches as self-hosted diabetes management agent

GlycemicGPT is an open-source, self-hosted platform that connects continuous glucose monitors, insulin pumps, and Nightscout instances to an AI analysis layer running on your own infrastructure. The creator is a Type 1 diabetic and software engineer who built it after going months without any clinician reviewing their data.

For anyone working on health-related agents, this is a concrete example of an end-to-end, devices plus data plus AI loop with real stakes. Self-hosting and local analysis will matter if you are dealing with HIPAA, EU health privacy, or very skeptical clinicians. It is still early and community validation, safety guardrails, and regulatory review are all open questions as of 2026-05-15.

Read more →

---

AWS details real-time multimodal voice agents with Nova 2 Sonic

Amazon Web Services describes how to combine Stream Vision Agents, Amazon Bedrock, and Amazon Nova 2 Sonic to build real-time voice agents that are “production-ready in minutes.” The post walks through the architecture, code samples, function calling, automatic reconnection, and multilingual voice support.

If you are building support or operations agents, this is essentially an opinionated reference implementation on top of managed services. You get low-latency streaming, tool use, and session resilience without owning the full media pipeline. The catch: you are tying your stack to Amazon Nova 2 Sonic and Amazon Bedrock pricing and limits, so cost and latency benchmarks in your environment really matter as of 2026-05-15.

Read more →

---

Amazon Lex introduces Assisted NLU to improve bot accuracy

Amazon Lex Assisted natural language understanding (NLU) is a new capability that helps you design and validate intents and slots so bots recognize user input more accurately. The blog shows how to write better intent and slot descriptions, use Test Workbench to validate performance, and migrate from traditional NLU to Assisted NLU.

For teams running contact center bots or internal helpdesk agents, this is about moving some of the “prompt engineering” mindset into classic NLU systems. You keep your Lex investment, but get guardrailed assistive tooling instead of brittle intent lists hand-tuned by one expert. It is still early and there are no independent benchmarks, so you will need to A/B against your existing bots as of 2026-05-15.

Read more →

---

Quick Hits

• Nokia embeds agentic AI across its fixed network platforms Nokia is rolling out agentic AI across broadband infrastructure so field techs and helpdesks can diagnose and fix network issues before customers notice. If you work on telco or large fleet operations, this is a signal that agentic workflows are moving into conservative infra.

• Work with Codex from anywhere OpenAI now lets you monitor and steer Codex coding agents from the ChatGPT mobile app, including approving commands and switching models from iOS or Android. Also covered by: TechCrunch: OpenAI says Codex is coming to your phone

• Helping ChatGPT better recognize context in sensitive conversations OpenAI describes new safety updates so ChatGPT can track risk over longer conversations and respond more appropriately. If you deploy agents in healthcare, finance, or youth-facing settings, these patterns are worth studying as of 2026-05-15.

• Sea's view on the future of agentic software development with Codex Sea Limited explains how it is rolling out Codex across engineering teams to standardize AI-native development. Good case study if you are trying to scale coding agents in a large org, especially in Asia-Pacific.

• Practice of law: case studies The Financial Times highlights PwC Singapore and Pro Bono SG using retrieval-augmented generation (RAG) chatbots and agents to support legal aid. Useful data point for anyone building legal or compliance-focused assistants.

• Granite Embedding Multilingual R2 IBM Granite releases two Apache 2.0 multilingual embedding models on ModernBERT, including a 97M parameter model claiming best sub 100M retrieval quality with 32K context. Strong option if you need open multilingual retrieval as of 2026-05-15.

• Unlocking asynchronicity in continuous batching Hugging Face details how separating CPU and GPU work in continuous batching can significantly boost large language model inference throughput. If you run your own serving stack, this is a practical guide to squeezing more QPS out of existing GPUs.

• Text analysis for hybrid search Weaviate breaks down tokenization, stopwords, and accent folding for hybrid search plus a /v1/tokenize endpoint. Worth a skim if your RAG results suffer from multilingual or accent-heavy queries.

• Control where your AI agents can browse with Chrome enterprise policies AWS shows how to lock Bedrock AgentCore browser agents to specific domains using Chrome enterprise policies and custom root CAs. Good pattern for regulated environments that want browsing but need strict controls.

• Launch HN: Ardent – Postgres sandboxes in seconds Ardent spins up production-like Postgres sandboxes with zero migration so humans and coding agents can test without touching real databases. If your agents write SQL, this sort of isolation is quickly becoming table stakes.

• PyTorch simulator refutes an 18-year-old quantum theorem A GitHub repo claims to refute a long standing quantum information result using a PyTorch based simulator. Very niche, but an interesting example of ML tooling being used as a scientific computation environment as of 2026-05-15.