How One Team Built a Vendor-Free On-Premise Data Lakehouse Using Open-Source Tools

A development team has detailed how they built a fully on-premise Data Lakehouse without proprietary software or cloud dependency, addressing budget and compliance constraints. The stack combines MinIO for storage, Apache Iceberg as the table format, Project Nessie for metadata cataloging, and Trino as the SQL engine, running on bare metal servers alongside Docker-hosted support services. The architecture follows a three-tier Medallion model — Bronze, Silver, and Gold layers — with governance responsibilities split between IT and business-facing teams like QA and BI. Pipeline orchestration is handled by Dagster, while dlt and dbt manage data ingestion and transformation respectively. The team plans to evolve toward real-time data ingestion in a future version by introducing Change Data Capture via Debezium and Kafka, with future posts planned on securing AI-generated access to the platform.

RS485 Fixes JK BMS–Deye Inverter Link After CAN Bus Fails in Off-Grid Solar Setup

A DIY off-grid solar builder successfully established closed-loop communication between a JK BMS and a Deye SUN-6K-SG04LP1 inverter after the commonly recommended CAN bus connection repeatedly failed to produce a stable handshake. The working solution used RS485 via the BMS's UART1 port, the PYLON protocol (014) on the inverter side, and a baud rate of 9600. Once configured correctly, the inverter began reading real-time battery data — including state of charge, pack voltage, and charge/discharge limits — directly from the BMS instead of estimating from voltage alone. The author notes that LiFePO4 batteries have a notoriously flat voltage curve, making voltage-only mode unreliable for accurate charge management. Key troubleshooting lessons include matching baud rates exactly, selecting the correct UART port, and verifying the inverter's protocol number, as mismatches produce silent failures indistinguishable from wiring faults.

Nylas enables AI agents to autonomously RSVP to calendar meeting invites

Nylas, a communications API platform, has introduced a workflow that allows AI agents with their own email addresses to respond to meeting invitations like a human attendee. When an invite lands in an agent's mailbox, Nylas automatically parses the iCalendar attachment and creates a corresponding event on the agent's calendar. The agent is listed as a participant with a pending status, and a webhook triggers the agent's code to check availability and fire an RSVP call. This approach positions the AI as an invitee rather than an organizer, meaning it can accept, decline, or tentatively accept meetings but cannot modify or cancel them. The feature relies on Nylas's Agent Accounts, which provide a real mailbox and calendar on hosted infrastructure, using the same API endpoints as any other connected account.

Developer halves $312 Anthropic bill by adding two bash hooks to Claude Code

A developer cut their monthly Anthropic API bill from $312 to $156 by implementing two bash hooks in Claude Code that prevent costly token-burning retry loops. Around 60% of the original bill stemmed from Claude repeatedly attempting to diagnose failed Cloudflare D1 database migrations, with each retry spiral consuming 40–60K tokens. A PostToolUse hook now intercepts failed migration commands immediately and flags a Cloudflare KV key, triggering a Slack alert so the developer can intervene before the agent investigates further. A second PreToolUse hook blocks deployment commands when the agent is operating on the main branch, a safeguard added after a misdirected production deploy briefly broke two Workers. The developer notes that hooks returning exit code 2 pass stderr messages directly to the agent as context, making specific, instructive error messages far more effective than generic block notices.

AI's Real Value Is Automation, Not Just Answering Questions

A growing gap exists between what AI systems can do and how most people actually use them, with many still treating powerful models as simple question-and-answer tools. Much of the modern workday is spent moving information rather than creating value — copying data, monitoring dashboards, and sending follow-ups — tasks that require persistence rather than creativity. Unlike humans, software does not get distracted, bored, or need sleep, making it far better suited for repetitive, continuous coordination work. The article argues that AI's transformative potential lies not in the model itself but in autonomous, ongoing workflows that operate without human supervision. Just as factories and offices were reshaped by machines and computers eliminating prior bottlenecks, AI is now replacing the coordination bottleneck — and human speed is increasingly what slows systems down.

Why Production AI Agents Like Hermes Skip LangChain for Hand-Rolled Loops

Many AI agent projects, including Hermes, OpenHands, Aider, and Codex CLI, opt to build custom orchestration loops instead of relying on frameworks like LangChain or LangGraph. The core agent loop is relatively simple, but production requirements such as streaming, budget control, provider failover, and prompt caching demand precise control that generic abstractions often obstruct. Framework-unified LLM interfaces also tend to lag behind vendor-released features, leaving teams unable to adopt new capabilities immediately. Hand-rolled code offers cleaner stack traces, better readability, and easier debugging compared to multi-layer framework abstractions with implicit behaviors. Hermes further reflects this philosophy through strict dependency pinning and lazy-loaded provider packages, explicitly minimizing supply-chain risk following a 2026 package-based cyberattack.

Cloudflare Durable Objects fix stops 800 Workers hammering origin on cache miss

A developer running a Korean flash sale was hit with an unexpected $40 bill after 800 concurrent Cloudflare Workers simultaneously missed a KV cache and flooded the origin server with requests within a 120ms window. The root cause was a timing gap between KV write latency (~80ms) and origin fetch time (~120ms), allowing hundreds of Workers to independently treat the cache as cold. The solution, known as request coalescing, ensures only one Worker fetches from the origin while all others await the same in-flight promise. This was implemented using a Cloudflare Durable Object, which runs in a single-threaded environment, making it safe to store and share in-flight promises without race conditions. A key refinement was assigning one Durable Object instance per resource key rather than a single global instance, preventing slow fetches for one resource from blocking others.

Google launches Agent-Ready Toolkit for Chrome to help sites work with AI agents

Google's Chrome team released the Agent-Ready Toolkit on June 22, offering developers a first-party set of tools to check and improve how websites behave when accessed by AI agents rather than human users. The toolkit includes a new Agentic Browsing category in Lighthouse, Chrome DevTools for Agents, Modern Web Guidance, and WebMCP, which lets sites expose their actions and forms to AI agents in a structured way. The tools perform static audits, verifying whether a site has properly registered tools, flagged unexposed forms, and validated input/output schemas. However, the toolkit cannot track whether a real AI agent has actually invoked those tools in production, leaving a gap between exposure and real-world usage. Google has confirmed that Gemini in Chrome will soon support WebMCP APIs, making runtime instrumentation of tool calls an increasingly important next step for developers.

Ireland Stun Pakistan in Historic Test Series Win at Home

Ireland claimed a landmark Test series victory against Pakistan in what players are calling the biggest achievement in the team's history. Batter Harry Tector described the week as the greatest in Irish cricket, regardless of how individual results are ranked historically. The series win marks a significant milestone for a nation that only entered Test cricket in 2018. Ireland's players and fans celebrated the result as a defining moment that signals the program's growing stature in world cricket. The victory is expected to boost confidence and investment in the country's Test cricket ambitions going forward.

ET Wealth | Avoid 7 ITR filing common mistakes

Own foreign assets? File your ITR the right way

Can mutual fund SWPs safely pay your home loan EMI?

Unclaimed insurance money? Check if it's yours

EPFO 3.0: Can it finally fix PF hassles for members?

NSE IPO: What investors should know before buying

Using UPI? Avoid mistakes with these 7 tips

Money lessons for children: How parents can start

Global equity returns: Which market performed best?

EB-5 Investor Visa guide for US Green Card

Four Parallel AI Agents Cut 300 i18n Bug Candidates Down to 60 Real Leaks

A software team tackling hardcoded Japanese text in a bilingual app deployed four parallel AI investigation agents, each assigned a different codebase surface area, to identify internationalisation leaks. A repository-wide grep initially returned thousands of hits, but the agents narrowed these down to roughly 300 candidates. To eliminate false positives, developers wrote a Python script using the AST module to detect whether Japanese string literals were already wrapped in language-conditional branches, reducing the list to about 60 genuine bugs. Among the confirmed leaks was a significant issue where all four Stripe webhook emails — including purchase confirmations and payment failure notices — were hardcoded in Japanese, meaning English-paying users had been receiving them in the wrong language for months. The exercise demonstrated how combining parallel AI triage with automated static analysis can make large-scale localisation audits tractable.