|
WildEdge
Live production monitoring
|
Google AI Edge Portal
Pre-production benchmarking
|
|
|---|---|---|
| Stage | Development and production, on the same instrumentation |
Pre-production only |
| Devices | The real devices your users hold |
Google’s lab fleet, 100+ Android device models |
| Models | Custom & trained models, self-hosted LLMs, remote LLMs, agentic workflows |
LiteRT only |
| Platforms | Python, Android, iOS |
Android |
| Measures | Latency, confidence drift, error rates, thermal state and CPU throttling, broken down by device cohort |
Latency, peak memory, initialization time, accelerator allocation |
| Coding agent access | An MCP server with scoped tokens and an audit trail. Ask your coding agent why a model regressed and it queries the events itself |
None. Benchmark results are read in the web console |
| What leaves your control | Structured telemetry only. Raw inputs stay on the device by default |
Your model, uploaded to Google’s cloud |
| Availability | Available now, instrument in minutes |
Private preview, allowlisted Google Cloud customers |
A lab fleet never produces the cases your model gets wrong. Real users do, and those are the ones worth training on.
Filter production events to low confidence, user corrections, or a model disagreeing with the one that took over from it.
Override what the model predicted, in place, keeping the record of what was changed and by whom.
Integrate your training pipelines against a manifest naming an immutable set of records. The dataset and the model stay reproducible.
Add the SDK and watch real inference across the devices, runtimes and providers you already use.
Competitor details reflect their publicly documented capabilities at the time of writing. If something here is out of date, tell us at [email protected] and we will correct it.