I really enjoyed reading your blog post, it was pretty informative about the techniques. I had built mine as a prototype and specifically interested in running all models local and in browser.
Tldraw seems to be a better medium for this that mermaid.
Would be curious to connect and compare notes! Yours is probably way better, I had built mine as a prototype and specifically interested in running all models local and in browser.
Cool to see that you did transcription AND LLM for the diagram with "in-browser" / local.
Tangent: But I'm wondering whether there would be something from the browser side to enforce this?
Like a sandbox so that my browser allows me to launch this via "download" only so that after I download everything there is no way for the website to get any of my data.
As a (non-technical) user what kind of guarantee do I have that the prompts, transcript etc. are treated only local?
Shimmer transcribes your meetings and generates live Mermaid diagrams as the conversation unfolds. Speech recognition (Whisper or Voxtral) and diagram generation (Qwen 3.5) both run on-device via transformers.js and WebGPU. No audio leaves your browser.
You talk, it listens, and a diagram evolves in real-time. It auto-detects the best diagram type (flowchart, sequence, mindmap, timeline, ER, state, class) or you can pick one manually. A 3-step LLM pipeline runs every minute to classify the conversation, extract structure, and generate Mermaid daigrams, building on the previous version each cycle.
You can also bring your own provider and API key (OpenAI, Anthropic, Ollama) for higher quality diagram generation while keeping transcription fully local.
I built this because I found myself frequently wanting to diagram what we're talking about in work meetings at my day job. Please give it a look and let me know what you think!
P.S. The name is inspired by "shimmer" being a collective noun for a group of mermaids.
I've been closely following the Astral team's work. I am excited by this release and look forward to trying out uv again.
I have been working with python for over 10 years and have standardized my workflow to only use pip & setuptools for all dependency management and packaging [1]. It works great as a vanilla setup and is 100% standards based. I think uv and similar tools mostly shine when you have massive codebases.
This is pretty cool and I'm surprised it hasn't gotten more traction on HN. The value in this for me will be whether the curated lists contain good content. I've started one track and am enjoying the content so far.
Their homepage says "Norse Tracks over 200,000 tor exit nodes". Tor metrics [1] says there exist 1,000 ish Tor exit nodes. So is Norse's statement a blatant lie?
I really enjoyed reading your blog post, it was pretty informative about the techniques. I had built mine as a prototype and specifically interested in running all models local and in browser.
Tldraw seems to be a better medium for this that mermaid.