Browser-based skyglow sim uses real photometric data

TLDR: A WebAssembly skyglow sim for real streetlights, a local AI data analyst that spits out notebooks, and a privacy-first PDF copilot all quietly level up your tooling.

Browser-based light pollution sim renders real streetlights and skyglow

The iesna.eu Skyglow Analysis demo runs entirely in the browser and uses real luminaire photometric files to simulate streetlight behavior and skyglow in urban scenes. It parses industry formats like LDT, EULUMDAT, and IES LM-63, then feeds that into a Bevy-based renderer that visualizes how different fixtures and layouts affect light spill and the night sky.

For anyone building simulation-heavy agents or planning tools, this is a clean example of serious physics-based computation delivered via WebAssembly with no backend. The author also supports design calculations against standards such as EN 13201 and ANSI / IES RP-8, so the numbers are anchored in real regulation rather than toy examples.

The big picture: a lot of agent use cases around planning, zoning, and urban design will need this level of fidelity and client-side performance as of 2026-05-02.

Read more →

---

MLJAR Studio turns natural language data chats into notebooks

MLJAR Studio is a desktop app that lets you query tabular data in natural language while running all analysis locally, then saves the full interaction as a reproducible Jupyter notebook. It wraps the open source mljar-supervised AutoML library, auto-manages a Python environment, generates code, executes it, and keeps the code and outputs aligned with the conversation.

This is aimed squarely at data scientists and analysts who like chat-based workflows but do not trust black-box SaaS tools or ephemeral GUIs. You get a local agent that operates on your files, with artifact ownership baked in: the result is an .ipynb you can commit, diff, and rerun. That is a big step up from "ask your data" demos that leave nothing audit-able.

For people building internal AI agents, the pattern is the lesson here as of 2026-05-02: pair a conversational layer with deterministic code generation and persistent artifacts.

Read more →

---

SimplePDF Copilot uses client-side tools to edit and fill PDFs

SimplePDF Copilot is an AI assistant that can fill form fields, answer questions, focus specific fields, add fields, and even delete pages within a PDF editor, while keeping the PDF itself fully client-side. Parsing, rendering, and field detection all run in the browser, continuing SimplePDF’s long standing privacy-respecting model that already serves more than 200,000 monthly users.

For workflow and document automation engineers, the interesting part is how the agent calls client-side tools instead of shipping documents to a server. Only the text strictly needed for language model reasoning is sent to the backend, which lets enterprises keep sensitive documents local while still getting AI assistance.

As of 2026-05-02 this is a concrete example of agentic tool use inside a mature product, not another greenfield demo.

Read more →

---

Quick Hits

• SAS: Using AI Agents to Revolutionise Fraud Detection - FinTech Magazine SAS is pitching new AI agents for fraud detection and supply chain optimization that expose a chat interface and some explainability hooks; mostly vision and positioning, with few hard benchmarks as of 2026-05-02.

• Autonomous SOC: The Evolution of Self-Driving Security Operations - Security Boulevard A good overview of "autonomous SOC" patterns that blend runbooks, structured automation, and agentic AI for high volume security workflows; useful if you are mapping which incidents can safely be delegated to agents first.

• Cybersecurity Agencies Worldwide Warn About Agentic AI Risks - Bloomberg Law News Cybersecurity agencies in the United States, United Kingdom, and Australia jointly warn that autonomous AI agents can be misused or breached, and urge companies to do explicit risk assessment and operational planning before deployment.

• Configuring Amazon Bedrock AgentCore Gateway for secure access to private resources AWS walks through using Amazon Bedrock AgentCore Gateway and Resource Gateway ENIs to let agents call private APIs, MCP servers on Amazon EKS, and other VPC resources without punching broad holes in your network.

• Reinforcement fine-tuning with LLM-as-a-judge Amazon details reinforcement learning from AI feedback (RLAIF) for Nova models using an LLM-as-a-judge setup, useful if you are experimenting with preference optimization or custom reward shaping.

• We need RSS for sharing abundant vibe-coded apps Simon Willison argues for RSS style feeds for micro-apps produced by "vibe coding", framing apps as frequent, personal posts that need standard distribution and install targets.

• [AINews] Agents for Everything Else: Codex for Knowledge Work, Claude for Creative Work Latent Space reflects on coding agents "breaking containment" and sketches a split between specialized code agents and more open-ended creative agents for the rest of knowledge work.