Free
$0
No credit card required
- 3 projects
- 100k events / month
- 14-day telemetry history
- 250 MB of input/output uploads / month
- 14-day captured-file retention
- Dataset creation, label review & export
For AI running on devices and in the cloud.
Know where your AI struggles in the real world.
See where model behavior changes across releases, runtimes, and devices. Inspect captured inputs, outputs, and feedback, then turn selected examples into datasets for evaluation and training.
Input/output capture is optional · No credit card required
Add WildEdge to this project. Identify whether it uses Python, Android, iOS, or macOS, install the appropriate WildEdge SDK, and instrument its local model runtimes and cloud model API calls. Keep raw input and output capture disabled by default. Use the WildEdge integration documentation: https://docs.wildedge.dev/
Pick a modality
Scan productCloud fallback
Product not recognized44% confidence
product-scanner · v4.2Production
Cloud fallback6%up 4 pts
Fleet confidence0.87down 0.03Fleet avg
Runtime errors2.1%no change
p95 latency340msno change
Product recognition dropped on one iPhone cohort.
Mean prediction confidence · daily7 days agomodel v4.2 rollouttoday
iPhone 14 · Neural Engine 1,240 scans0.44
iPhone 14 · CPU 9,860 scans0.90
Confidence by class: bottle 0.38 boxed item 0.41
Observed after the model rollout · cause not yet confirmedInspect affected devices ↓
Task completion82%down 9 pts
Correction rate14%up 8 pts
Fallback rate18%up 11 pts
p95 latency680msno change
Users are correcting more names and numbers.
One corrected interactionAssistant heard“Call Ann”
User corrected“Call Anne-Marie”
Failed stageContact resolutionTranscript accepted; no matching contact was selected
Correction rate after Android build UP1A.231005: 14%, up from 6% across 6,140 interactions.
860 corrected interactions available for review · evidence capture enabledReview corrected interactions ↓
Success after retry91%down 2 ptsRecovered
First-response validity84%down 13 pts
Retry rate16%up 13 pts
p95 latency1.2sno change
A new prompt is producing invalid fields.
Prompt v12 · small-model-v3
Model output
{"priority": "urgent", "category": "billing"}
Schema rejected
priorityExpected low | medium | highFirst-response validity: 71% for prompt v12 + small-model-v3, versus 97% for prompt v11 on the same model.
2,180 affected runs · recovered responses include one retryInspect validation failures ↓
Works with
Core MLMLXTFLiteONNX RuntimeML KitExecuTorchllama.cpp
OpenAIAnthropicGemini
See the product
From a bad prediction to a better model.
WildEdge flags models that drift from their own history. This one lost 17% confidence after an app update, with no crash and no error.
View full screenshot ↗
Filter by device, app version, or runtime. On iPhone 14, confidence fell 37% the day version 1.38 rolled out.
View full screenshot ↗
Open a low-confidence prediction to see the input. This BMW was classified as a Ford, at 33% confidence.
View full screenshot ↗
Collect the misreads into a dataset, check them against user corrections, and export them for retraining.
Integrations
Connect your models.
Add an SDK to collect inference telemetry. Enable input and output capture when you need examples for review.
Python
uv add wildedge-sdkTransformers · PyTorch · TensorFlow · Keras · ONNX · MLX · llama.cpp · Ultralytics · timm · OpenAI · Anthropic · OpenAI-compatible APIs
Android
implementation("dev.wildedge:wildedge-android:0.1.0")TFLite · Play Services TFLite · ONNX Runtime · ML Kit · LiteRT · LiteRT LLM · Gemini · Custom models
iOS & macOS
.package(url: "https://github.com/wild-edge/wildedge-swift.git", from: "1.1.0")Core ML · TFLite · ONNX Runtime · ML Kit · ExecuTorch · llama.cpp · OpenRouter · Gemini · Custom models
What happens on real devices
The same model behaves differently across devices.
Where your model actually runs
NNAPI GPU CPU
| Chipset | Execution path | p95 latency |
|---|---|---|
| Tensor G3Pixel 8, Pixel 8 Pro | 180ms | |
| Snapdragon 8 Gen 2Galaxy S23, OnePlus 11 | 210ms | |
| Tensor G2Pixel 7, Pixel 7a | 340ms | |
| Exynos 1380Galaxy A54, Galaxy M54 | 520ms |
The bottom two chipsets can run on NNAPI. Most of the time they fall back to CPU, and take nearly 3x longer.
Work with your coding agent
From production failure to the code behind it.
Give your coding agent access to WildEdge production telemetry. It can investigate affected devices, inspect traces, and connect runtime failures to your code.
Connect your coding agent Powered by WildEdge MCP› Our brand classifier looks worse lately. What changed, and where?
✓ Compared model versions by device
✓ Split A15 iPhones by accelerator
✓ Checked latency and user feedback
App 1.38 shipped effv2s-brand-s0 v12. The drop only happens on A15 iPhones running v12 on the Neural Engine.
| A15 confidence | v11 | v12 |
|---|---|---|
| Neural Engine | 0.77 | 0.31 |
| GPU | 0.76 | 0.77 |
Latency is unchanged, so the model runs normally and returns wrong answers. User feedback on these phones rose from 7% to 28% of scans.
Likely cause: a v12 layer that the A15 Neural Engine computes incorrectly in fp16. Quick fix: set computeUnits = .cpuAndGPU for v12 on iPhone14,x, then compare layer outputs on an A15 device.
Evidence: version × device × accelerator · feedback rate · latency
Device to cloud
Follow one request across device and cloud.
When a local model hands work to a cloud model, inspect both steps in the same trace. See where time was spent and connect the response to the recorded user outcome.
Compare WildEdge with other toolsExample trace
tr_01J8Y2Q41.04 s total- Device requestiPhone 14 · iOS 18.1received
- Core ML inferenceproduct-scanner v4.2 · Neural Engine0.44 confidence
- Cloud fallbackcloud-classifier-v3 · eu-central842 ms
- User outcomesuggestion shown in applicationaccepted
Model v4.2Raw input not retainedReview-ready
Private deployment
Deploy WildEdge where your data lives.
Run WildEdge on-premises or in your own cloud. Structured records use an open table format, with captured inputs and outputs in object storage.
Input/output capture is optional
On-premises deployment
Bring your own cloud
Explore compliance evidence ↗
Pricing
Try WildEdge free. Choose a plan as your usage and deployment needs grow.
Pro
$299 / month
Billed monthly
- Unlimited projects
- 1M events / month
- 90-day telemetry history
- 10 GB of input/output uploads / month
- 90-day captured-file retention
- Dataset creation, label review & export
- Email support
Enterprise
Let’s talk
For private deployments and scale
- Everything in Pro
- VPC, on-premises & air-gapped options
- Your own object storage
- Custom retention, SSO & audit logs
- Dedicated support & SLA
Need more upload capacity? Talk to us about additional capacity ↗
Event and upload allowances are shared across your company’s projects and reset each monthly usage period. Input and output files both count toward the upload allowance. Adding a file to a dataset does not extend its retention; export files you need to keep before they expire. Training integrations require an account with the destination provider.