Drop a TFLite or ONNX model
Parse the model locally, inspect graph structure, and prepare on-device optimization evidence.
Local-only
Model weights stay in browser memory.
No upload
The selected model is not uploaded by this app.
No telemetry
No model, tensor, input, output, or report logging.
Local execution
Analysis runs locally on the user's device.
On-device Audit Dashboard
Parsed TFLite/ONNX facts plus static heuristics for cache pressure, roofline posture, graph risk, and local execution readiness.
--
Waiting for model
Load a .tflite or .onnx file to generate an optimization profile.
Static Roofline Mix
Top Signals
static inspection queuePriority Actions
model-specificLocal Preprocessing Pipeline
Prepare a real image tensor locally before running browser-side inference or benchmark checks.
Inference Benchmark
Runs fake-input inference locally. The model is still not uploaded.
| Backend | Compile | p50 | p90 | p99 | Mean | Status |
|---|
Graph Explorer
Search ops, inspect tensor metadata, and trace local producer/consumer links.
| # | Op | Out | MACs | Bound | XNNPACK | Quant | Align | Pack | Fusion | Overlay |
|---|
Stage View
Operator Histogram
| Op | Count |
|---|
Top MAC Ops
| # | Op | Output | MACs | Bytes |
|---|
Static Roofline
| # | Op | Intensity | Guess | Action |
|---|