Article URL: https://simonwillison.net/2026/Aug/16/qwen-38-27b/ Comments URL: https://news.ycombinator.com/item?id=49324985 Points: 17 # Comments: 3
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
A developer has built and deployed a production-ready chat application from scratch, documenting the full process publicly. The app uses Node.js, Express, Socket.io, and MongoDB on the backend, with Vanilla JavaScript handling the frontend. It features JWT-based user authentication, real-time messaging via WebSockets, and automatic message deletion after 30 days. The application was deployed on Render.com's free tier by connecting a GitHub repository, going live within minutes. Key challenges during development included resolving MongoDB DNS issues, configuring Socket.io CORS settings, and managing socket connections gracefully.
A developer has launched PawServe, a progressive web app designed to consolidate dog-care resources into a single platform. The app allows users to search and filter over 20 dog breeds by size, temperament, and category, and browse a directory of veterinary specialists across multiple fields. An AI-powered symptom checker, built using Google Gemini, lets owners describe their dog's condition via text or voice and receive first-aid-style guidance with severity cues. Sensitive API calls are routed through Supabase edge functions to keep credentials secure on the backend. Additional features include a pet supply shop, dog-friendly venue listings, training schools, parks, and care guides, all accessible through a React and TypeScript-based interface deployed on GitHub Pages.
A developer found that loading Whisper, bge-m3, and Gemma simultaneously caused out-of-memory errors on their workstation, which had enough VRAM for only one model at a time. Rather than buying new hardware or splitting workloads across machines, they adopted a sequential loading strategy — loading, using, and unloading each model before moving to the next. The approach works because the tasks do not overlap in time, and the added load latency of a few seconds per model is negligible against longer processing times like a 90-second transcription. Model selection involved trade-offs: bge-m3 was chosen over a smaller embedding model for better Korean-English cross-lingual accuracy, and Gemma edged out LLaVA variants in informal document-image description tests. The setup has run stably for several weeks, and the developer is now working to connect the three models into a unified pipeline for processing calls, emails, and messages.
A developer discovered that backing up a Supabase project involves far more than a simple database dump and storage copy, after a Reddit discussion highlighted critical gaps in common backup approaches. Issues such as non-atomic snapshots, missing auth data, storage metadata, and unverified restorability prompted them to build a dedicated tool called pgDumpster. The CLI handles PostgreSQL roles, schema, and data while generating a portable archive with a manifest, integrity details, and explicit coverage limitations. Rather than claiming complete backups, pgDumpster surfaces what cannot be captured or restored identically, making gaps visible instead of silently ignoring them. Released as version 0.1.2, the tool is publicly available on GitHub and npm under a PolyForm Shield licence, free for personal and self-hosted use.
Discussion (0)
Log in to join the discussion and vote.
Log in