Back to feed
Dev.to
Dev.to
7/24/2026
Building browser-only AI tools: debugging broken ONNX exports and ONNX Runtime fusion bugs

Building browser-only AI tools: debugging broken ONNX exports and ONNX Runtime fusion bugs

Original: Context Compression: Making AI Agents Forget Without Losing the Plot

Short summary

A developer shipped three browser-only AI tools (background removal, stem separation, speech-to-text) and documents what broke. A broken HTDemucs ONNX export produced physically impossible output (stems at 145% mix magnitude), solved by exporting from the official checkpoint. An ONNX Runtime fusion bug crashed Whisper due to tied embeddings, requiring a workaround. The post is a detailed debugging log with concrete data and hard-won lessons for on-device AI.

  • •A broken ONNX export of HTDemucs produced physically impossible stem magnitudes, proven by comparing against the PyTorch reference
  • •ONNX Runtime's graph optimizer crashes on Whisper's tied embeddings during quantized session creation
  • •All three tools run entirely client-side with no file uploads, using transformers.js and onnxruntime-web

Generated with AI, which can make mistakes.

#ai-tools#open-source#industry-adoption
Read full article at Dev.to

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

The original title is 11 words: "Context rot: why your AI agent gets dumber the longer it runs"
Dev.to
Why Owning Your AI Agent's Context Is Crucial for Production
Dev.to
Treat the Context Window Like a Budget, Not a Junk Drawer
Dev.to
The original title is "ContextVault: Own Your AI Context Across Models, Agents, and Time"
Dev.to