First, a little context.
Your model’s receptive field sets the context size. While it fills, the stream reports “warming up.” No verdict yet means “not enough data,” not severity 0.
Have the whole signal already? Send it over.
Still collecting it? Send one value at a time. We’ll keep the context.
One request. A series of values. All the uh ohs we find.
Available with an API token ONGOING ANALYSISA rolling window. Fresh context. A verdict with each new sample.
Available with an API tokenFor a recorded run, a chunk of readings, or a signal you’ve already collected.
Send an array of values in one request. The detector moves through the signal, looking for moments that break the pattern.
The response includes the overall verdict, anomaly locations, scores, and emoji severities. No drama? 😌, severity 0, and an empty anomaly list.
Each request stands on its own. The local demo does not store your samples or remember the previous request.
POST /v1/analyze
Authorization: Bearer YOUR_API_TOKEN
Content-Type: application/json
{
"signal": [42, 41, 43, 42, 41, 43, 42, 85]
}{
"verdict": {
"emoji": "😱",
"severity": 8
},
"uhohs": [{
"index": 7,
"value": 85,
"emoji": "😱",
"severity": 8
}]
}For a sensor, a service metric, or anything with another reading on the way.
Instead of sending the entire history every time, start a stream and send each new value with its stream ID. We’ll keep a window’s worth of recent samples for that stream and use the updated window to analyze the newest value.
Send the next value to see what stays, what leaves, and what gets a verdict.
One more sample before the first verdict.
Illustration only: the 85 spike is scripted to show an uh oh. This does not run a model or store data on the server.
Your model’s receptive field sets the context size. While it fills, the stream reports “warming up.” No verdict yet means “not enough data,” not severity 0.
Once full, each arrival replaces the oldest value. A model keeps its receptive field’s worth of prior context to predict the next sample.
Each new value gets its own result based on the current window. Yesterday’s uh oh doesn’t make every future reading an uh oh.
Keep different sensors or metrics in separate streams. Send their samples in time order so the stored window represents the right history.
Start sequence 0 and use next_sequence from each response. Retry the last sequence with the same values to replay its result. Use a new stream ID to start fresh, or DELETE the old stream. Send samples in order; timestamps are not inferred.
POST /v1/streams/sensor-01/samples
Authorization: Bearer YOUR_API_TOKEN
Content-Type: application/json
{"sequence": 0, "value": 41}
Use next_sequence from the response.
Fill the model context, then get fresh verdicts.| One-time signal | Ongoing stream | |
|---|---|---|
| You send | A complete array of samples | A stream ID and each new value |
| We keep | No history between requests | A rolling window per stream |
| You get | A summary and anomalies across the signal | A verdict for each new value after warm-up |
| Best for | Recorded runs and collected readings | Live sensors and continuous metrics |
| Preview status | Token-based model inference | Persistent model history |
What’s running today? The homepage uses a statistical demo. Authenticated API calls run your saved TCN: complete signals at /v1/analyze, or ongoing samples at /v1/streams/{stream_id}/samples. Create a model-specific token in your workspace’s API tab. Stream history persists in the database. The five-sample illustration above is explanatory; real warm-up depends on your model.
Meet the faces before you meet the API.