I shipped three AI tools that run entirely in the browser — here's everything that broke

Short summary

A developer shares hard-won lessons from shipping three browser-only AI tools (background removal, stem separation, speech-to-text). Key issues included a broken third-party ONNX export of Meta's HTDemucs producing physically impossible output (stems at 145%+ of original magnitude), and an ONNX Runtime graph-optimizer bug that crashes on Whisper's tied embeddings. The article provides concrete debugging strategies for anyone building on-device AI features.

• •Broken ONNX model file detected via physically impossible output magnitudes (145%+ of input signal)
• •ONNX Runtime fusion pass creates MatMulNBits op at session creation, crashing on Whisper's tied embedding/lm_head pattern
• •Practical advice: validate model exports against reference implementations before shipping, suspect the file before the architecture

Generated with AI, which can make mistakes.