A self-tuning filter on the same stream: signal cleaning at three latency budgets — instant, low-latency, offline maximum quality — plus one-step prediction and gap recovery. No models to build, no filters to tune.
Pricing: Layer 1 — $29/mo; three output modes — real-time, small-lag, offline-grade · 5 channels free. Full pricing →
The problem
Most real-world signals are noisy, and classical tools force a choice: filters that are simple but fragile, or a hand-built model of the signal you rarely have time to construct — and have to redo for every new channel and noise level.
Three outputs
All three outputs run on the same stream. Most of the achievable denoising is bought by the first 80 samples of latency: moving from zero-lag to the low-latency output gains +6.1 dB on average — always positive across signal classes and SNRs (+0.8 to +14.7). Whole-signal context adds only +1.9 dB more on average.
| Output | Latency | What it sees | When to use it |
|---|---|---|---|
| instant | 0 · zero-lag | past + current only | hard real-time consumers that cannot wait a single sample |
| low-latency | 80 samples | past + 80 ahead | streaming pipelines — most of the quality at bounded delay |
| offline | whole signal | global context | batch re-processing and archival cleanup at maximum quality |
On non-stationary signals the streaming low-latency output even beats the offline batch (moving-spectrum chirp at 20 dB: +5.3 vs +0.0 dB) — per-block adaptation tracks a moving spectrum that a single global method choice cannot. How we measure →
Evidence
dB RMSE-reduction vs the noisy input; higher is better, negative means the filter added error. Classical baselines: Butterworth, Kalman-AR1, Savitzky-Golay, wavelet. On these hard signals every classical filter is negative at high SNR — the streaming output stays strongly positive.
| Hard signal · SNR | Best classical filter | Low-latency (80 samples) | Offline |
|---|---|---|---|
| Multi-tone vibration · 30 dB | −6.5 dB (others to −26.5) | +8.3 dB | +12.5 dB |
| Multi-tone vibration · 20 dB | −9.3 dB (others to −16.7) | +8.2 dB | +12.8 dB |
| Multi-tone vibration · 10 dB | −5.3 dB | +8.2 dB | +12.8 dB |
| Chirp (moving spectrum) · 20 dB | −11.7 dB (others to −20.1) | +5.3 dB | +0.0 dB |
| Chirp (moving spectrum) · 10 dB | −7.4 dB (others to −10.1) | +4.9 dB | +0.2 dB |
On simple stationary signals classical filters also work — the gap opens exactly where they break: multi-tone structure, moving spectra, nonlinear oscillations.
When there is nothing to remove, the low-latency and offline outputs leave a near-clean signal essentially untouched (RMSE at the input floor) — a principled record/replay criterion backs the denoiser off to the raw signal instead of over-smoothing it.
The same stream also serves one-step prediction and recovery of missing samples — for feature pipelines that need every tick filled, and for post-processing archives with dropouts.
Who it's for
Filtered features and a fast model-free baseline before — or instead of — building and maintaining your own denoising models.
Cleaner signatures from test rigs and precision instruments without hand-designing a filter per channel and re-tuning it per experiment.
Batch re-processing of accumulated data at offline-max quality: cleanup, gap recovery and one-step prediction on the archive you already have.
Pricing
The filter is one of the four API surfaces of the stream integration, so it's priced by the same usage meter of the layer. Free tier to verify on your own signals; self-serve usage-based after the founding period; custom filtering integrations are available as a privately contracted service.
A few channels and capped points — enough to check the three outputs against your own hard signals in the free tier.
Usage-based self-serve after the founding period: all four surfaces of the stream — anomalies, connections, passport and filter — on one meter, ~$0.02–0.05*/channel-hour as the initial range.
Custom filtering deployments and integrations — including private/on-prem — as a privately contracted deployment on the profile.
* indicative starting range, validated in pilots — not a published list price. See Pricing for the full model.
Founding release
Multi-tone vibration, moving spectra, nonlinear oscillations — the signals classical filters fail on are exactly where this filter earns its place. It runs on the same stream as anomaly detection and connection discovery, so cleaning is rarely the only thing you'll want from that data.