Optimization acceptance
Accuracy gated
Candidates are checked against accuracy constraints. No device-specific speedup or NPU run is established by the preview.
Validation notes Back to portfolio
Current model and OOD controls with pipeline stages pending. No NPU measurements or completed run are shown.Open full capture
Case study / Optimization workbench
QualEdge
Compares quantization and execution-provider options while retaining accuracy gates, benchmark context, and fallback behavior.
Optimization workbench
Solo project; AI-assisted development.
The idea
The problem
A faster model is not useful when accuracy regresses or the preferred execution provider is unavailable.
The implementation
What I built
Built ONNX candidate comparisons with accuracy gates and explicit provider fallback.
- Python
- ONNX Runtime
- React
- Quantization
Features
- Compare ONNX optimization candidates
- Reject accuracy regressions
- Inspect execution-provider fallback
Architecture
- ONNX optimization and quantization pipeline.
- Provider benchmarking and accuracy comparisons.
- Browser workbench for results, rejected candidates, and fallback state.
Engineering decisions
- Check accuracy before accepting an optimization.
- Preserve the measured runtime and device context.
Validation and measured results
- Local workbench supports benchmark comparisons and accuracy-regression rejection.
- Hardware-specific results require actual device receipts; no blanket NPU speedup is claimed.
Limits and unfinished work
- Results are specific to model, data, provider, and device.
- Provider availability does not prove accelerator execution on an unmeasured device.
Evidence and sources
- Workbench validation notes
Locally reviewed implementation and experiment context; no universal hardware claim. The repository was not publicly accessible on 14 October 2026.