Field Notes

Better voice output, knowledge assistant without RAG ops, interactive charts right in chat

Three levers for your day: upgrade voices, use knowledge without running your own stack, make visualizations clickable.

Key AI news

Amazon Polly lifts voice quality without code changes The generative Polly engine delivers more natural voices with lower latency. 14 new voices in 9 language regions have been added, including Hindi and Finnish for the first time in general availability. In total, 57 generative voices across 11 languages and 25 locales are available. The switch is a drop-in with no new APIs. Price stays the same according to the vendor, and GovCloud US West is now included. If you run a contact center or voice agents, you get more accurate number, address, and code prompts with no deployment effort. Vendor claims.1

SharePoint knowledge without your own RAG infrastructure: IskayPet uses Bedrock Managed Knowledge Bases IskayPet built a multilingual internal assistant on Amazon Bedrock Managed Knowledge Bases. The solution pulls content directly from existing SharePoint via native integration, uses hybrid search with BM25 and semantic search, ranks with reranking, and cites sources. Access rights are passed through at document level so users only see what they are allowed to read anyway. Built with Kiro and Claude Haiku 4.5, running in production across four languages. If you want to consolidate scattered SharePoint knowledge, this saves you vector stores, embeddings, and your own ingestion jobs.2

Flint-interactive: clickable visualizations from chat and in the canvas Flint-interactive is a language based on Flint that describes interactions like zooming, filtering, linked brushing, and cross-chart actions declaratively. A compiler translates user actions into semantic events and updates visualizations via operators. You can use it in chat via Flint-MCP or embed it as a canvas in apps to connect text, tables, and widgets. You define tailored interactions via API without renderer-specific mark manipulation. Claims about higher reliability and less debugging come from the vendor.3

Rescale builds a triage agent with AI-DLC in days instead of months Rescale automates classification and routing of nightly end-to-end test failures. The agent runs as a step in Argo Workflows on EKS, is triggered by Playwright failures, classifies with Claude Opus 4.8, pulls patterns from a knowledge base, reads logs via MCP, and produces a structured remediation report for the teams. Humans review and decide. According to the case study, a staging system was built in five days and went live a few weeks later. Useful if test triage is a bottleneck in your release process.4

Highlights for your daily work

  • You already use Amazon Polly. Then the upgrade gives you audibly more natural voices without changing a single line of code. Test prompts with order numbers and addresses before you roll them out in IVR flows. Price remains the same according to AWS.1

  • You want to keep processes, people, and countries separate. Bedrock Managed Knowledge Bases pull in SharePoint content, evaluate hybrid search, and respect permissions. Start with one area, like onboarding, then work your way to procurement and store operations.2

  • You create reports with LLM support. Let Flint-interactive generate interactive charts right in chat and link them to text or tables. Ask the system for a linked selection between two charts to check patterns faster.3

Tools and updates

  • TorchNeuron for Trainium: open-source backend that runs PyTorch code on AWS Trainium with minimal changes. Adaptive eager mode and caching shorten startup and increase efficiency. Requirement: access to Trainium instances.5

  • Sakana Aquarium announced: platform for decision agents with an ontology layer, managed agents for workflows, and governance with traceability. Announcement, no GA details. Relevant if you want to build compliant agents on existing data.6

  • IBM Granite Time-Series Ensemble: ensemble forecasting for time series on Hugging Face. For demand, inventory, or load forecasting in planning and operations. Requires time series setup and validation.7

  • n8n template: AI videos with a consistent character from five photos. For brand presence and product clips if you work in n8n.8

Try it in five minutes

Your own practice idea: recreate cited answering like the IskayPet example without infrastructure

  1. Paste this sample text into your existing AI chat: "Return policy: customers can return goods within 30 days with receipt. Exceptions: live animals, refrigerated goods. Onboarding: new store employees receive safety training and cash register training on day one." Set the task: "Answer only from the text. Quote the relevant passage verbatim in quotation marks. If something is missing, write: Not in the documents."2

  2. Ask: "How long is the return period and are there exceptions?" Expect an answer with a correct quote and the exceptions listed. Then ask: "How do I apply for educational leave?" Expect: "Not in the documents."2

  3. Switch the language: "Responda em português com citação." Check if quotes and statement stay the same. Use the pattern later for your real policy texts.2

Result: the chat provides justified, cited answers or refuses cleanly. Checkpoint: does each answer contain a correct quote from the text or the exact sentence "Not in the documents".

Sources

  1. Amazon Polly delivers a major upgrade to its Generative engine (aws.amazon.com)
  2. How IskayPet unified corporate knowledge with Amazon Bedrock Managed Knowledge Bases (aws.amazon.com)
  3. Flint-interactive: A language for creating interactive data visualizations (microsoft.com)
  4. How Rescale built an AI triage agent with Strands Agents and AI-DLC (aws.amazon.com)
  5. The science behind TorchNeuron, Amazon’s new PyTorch backend for AWS Trainium chips (amazon.science)
  6. @sakanaailabs: 意思決定支援・実行のためのAIエージェントプラットフォーム「Sakana Aquarium」を発表

https://t.co/ZITxgmsV2T

Aquariumは、Sakana AIがこれまでのAI実装の現場で繰り返し直面してきた課題から生まれたプラットフォームです。

AIによる意思決定支援を実務につなげるには、モデルの性能だけでなく 、AIが組織 (x.com) 7. Ensemble forecasting with IBM Granite time-series models (huggingface.co) 8. @n8n_io: Featured template: make AI videos where your character looks the same in every shot. You upload 5 photos and the workflow does the rest. 👀 🔗 Check it out here: https://t.co/mNiQT8 (x.com)