GoodBarber used a local AI model to match 5,892 multilingual blog posts for hreflang tags

·3 views

GoodBarber's engineering team faced a 14-year backlog of 5,892 blog posts across seven languages with no system linking translated articles to each other. Without hreflang data, Google could not identify which posts were translations of the same content across subdomains like fr.goodbarber.com, www, es, it, pt, de, and nl. The team used two open-weight AI models — gemma2:27b and bge-m3 — running locally on a MacBook Pro with an M3 Max chip, without any cloud service. A Python script fetched all articles via the CMS API and used a date-window filter to narrow candidates before the language model matched translations by comparing article summaries. The project demonstrated that a two-year-old locally hosted AI model could handle a large-scale multilingual content matching task without external infrastructure.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Remap OpenCode Exit Key to Avoid Accidental Session Loss with Ctrl+C

OpenCode, a terminal-based AI coding tool, exits the entire application by default when users press Ctrl+C, a keybinding deeply ingrained in terminal workflows for interrupting processes. This behavior can cause frustrating loss of long AI sessions that may hold valuable context. The fix involves editing OpenCode's configuration file to reassign the app_exit action to a different shortcut, such as Ctrl+D or a leader-based key combination. This small change separates the 'interrupt current action' function from 'close the application', reducing the risk of accidental exits. Users who frequently work in terminals are advised to make this remapping early in their OpenCode setup.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Lightweight Angular Dashboard Using Signals and Tailwind CSS

A developer has released Exo-Dash, an open-source Angular dashboard starter kit designed to address the bloat and complexity common in legacy admin panel templates. The project replaces traditional RxJS-based UI state management with Angular Signals, enabling synchronous, readable state handling for elements like sidebars and dark-mode toggles. Built entirely with Standalone Components and no NgModule files, Exo-Dash is architected to align with Angular's push toward a zoneless future, resulting in a smaller initial bundle size. Tailwind CSS was chosen over component libraries like Angular Material or Bootstrap to avoid CSS encapsulation conflicts and to ship only the styles actually used. The project is positioned as a lightweight foundation for developers building SaaS platforms or internal tools who want a modern, clean starting point.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a B2B SaaS team used custom ESLint rules to enforce API access control

Engineers at Malibou, an HRIS and payroll platform built on Next.js, developed custom ESLint linter rules to enforce access control across all API routes and server actions. The core challenge was horizontal authorization — ensuring every database query validates that a specific row belongs to the requesting user's organization, not just that the user is authenticated. Because tenant-based row-level security in SQL could not capture complex, role-dependent business rules, the team chose to enforce their existing TypeScript logic structurally through linting. Their first custom rule requires every server entry point to pass through a central middleware that handles authentication and permission checks, causing non-compliant routes to fail CI. The approach addresses Broken Object Level Authorization, which has ranked at the top of the OWASP API Security Top 10 for several years.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Modern Cloud Has Seven Layers, Not Three — Here's What Actually Changed

The traditional three-tier cloud model of IaaS, PaaS, and SaaS no longer reflects how modern infrastructure actually works, with the stack having expanded to seven distinct levels over the past 18 months. The boundary between 'warm' always-on servers and 'cold' serverless functions has blurred, as platforms like Vercel and Google Cloud Run now offer hybrid billing models that combine traits of both. The term 'serverless' has also fragmented into at least four distinct categories — functions, containers, edge isolates, and serverless databases — each with very different constraints and limits. Meanwhile, significant capital has shifted toward the database layer, highlighted by Databricks acquiring serverless Postgres startup Neon for approximately $1 billion despite the company generating around $25 million in revenue. A new infrastructure layer has also emerged to support AI agents that autonomously write and execute code, signaling a fundamental reshaping of how cloud computing is structured and taught.

0 comments Read more at DEV Community