Domscribe vs Gemma 4 Multimodal Fine-Tuner

Domscribe is an open-source bundler plugin that solves a concrete, frustrating gap in AI-assisted frontend development: agents like Claude and Cursor are great at editing source files, but they have no way to trace which file owns a given rendered element. Domscribe assigns stable IDs to every DOM element at build time and generates a manifest mapping each element to its exact source file, component tree, props, and state. AI coding agents connect via MCP to query any live node in the browser — or click elements in a visual overlay to pass targeted UI context directly into the agent's tool call. The implementation is clean. All debug metadata is stripped at production build time, so there's zero runtime overhead. The manifest only ships in development, keeping bundle sizes clean. It supports React, Vue, Next.js, Nuxt, and all major bundlers: Vite, Webpack, and Turbopack. The MCP server can be pointed at any agent — Claude Code, Cursor, Windsurf, or raw Claude API via any compatible client. This is a genuinely practical tool for teams doing agentic UI work. The bidirectional bridge — source-to-DOM *and* DOM-to-source — means agents no longer need to guess which component renders what. It's MIT licensed, fully local, and has no cloud dependency. A small but meaningful infrastructure piece for the emerging agentic frontend workflow.

Gemma 4 Multimodal Fine-Tuner is an open-source toolkit that lets developers fine-tune Google's Gemma 4 and 3n models across all three modalities — text, images, and audio — using only Apple Silicon hardware. It runs natively on PyTorch with Metal Performance Shaders (MPS), bypassing the NVIDIA requirement that has historically blocked Mac users from serious local fine-tuning work. The toolkit handles the full training pipeline including dataset prep, LoRA adapters, and multi-modal data collation. It ships with working example notebooks, a validation suite, and clean abstractions that don't require deep familiarity with the underlying MPS stack. Apple Silicon's unified memory architecture actually helps here — large multimodal batches fit in memory that would otherwise require GPU VRAM splitting on CUDA setups. Posted to Hacker News on April 7 as a Show HN, it pulled 109 upvotes and 165 GitHub stars within hours. The timing is sharp: Gemma 4 just dropped days ago with new multimodal capabilities, and the community immediately wanted local fine-tuning. This fills that gap faster than Google's own tooling.