Compress sensor data to the rule that made it.
For industrial time-series, AT-1’s generative tier discovers the model behind a signal — a damped harmonic, an exponential decay, a linear recurrence, a sequence of regimes — and stores that model plus a small irreducible residual. The archive is byte-exact, provably never larger than xz, and queryable straight from the model.
- 1.3–1.5×
- smaller than xz on real industrial data
- byte-exact
- SHA-256 verified, every record
- never worse
- provably ≤ xz on any input
- query-from-model
- answers thresholds without decompressing
We discovered that most of what you store was never information in the first place. It was the inevitable output of rules you already knew. We’re giving you back the rules.
Three guarantees, not a black box
Byte-exact, or it doesn't ship
Each encode is re-decoded and byte-compared to the original at encode time, and carries an embedded SHA-256 the decoder re-checks. Mandatory for warranty forensics and regulated records.
Never larger than xz
A general-compressor baseline is always in the selection, so the result is provably never materially larger than xz/zstd on any input. A model is used only when it is strictly smaller.
Query straight from the model
Because the archive stores the generating model, threshold-crossing queries — “when does this vibration amplitude cross the alarm band” — are answered by inverting the model, with no residual decompression.
Validated on real public data
Every trace below is real, public, and reconstructed byte-for-byte. Ratios are vs xz (>1 = smaller than xz). We don’t overclaim: on broadband real sensor data the win is a consistent 1.3–1.5×; on clean, structured signals (steady vibration, multi-regime traces) it reaches 2–2.5×. The guarantee — byte-exact and never worse than xz — holds on all of them.
| Dataset | Vertical | vs xz |
|---|---|---|
| CWRU bearing vibration (4 fault states) | rotating machinery | 1.34–1.52× |
| NASA C-MAPSS turbofan sensors | gas turbine | 1.16–1.38× |
| Paderborn motor phase current (64 kHz) | rotating machinery | 1.33–1.47× |
| UCI hydraulic pressure (cyclic) | factory automation | 1.44× |
| NASA ICESat-2 ATL03 altimetry | satellite altimetry | 1.03–1.12× |
How a deployment works
The heavy encoder runs on-prem — your sensor data never has to leave your network — and the ~170-line decoder is free and runs everywhere, from edge controllers to the cloud.
Try
Run a sample of your sensor data and see the ratio, byte-exact check, and a live threshold query — nothing leaves your machine.
Pilot
We deploy the on-prem encoder against a historian export and report the ratio + byte-exact result on your tags.
Deploy
A connector tiers cold tags to .at1gen; the free decoder reads everywhere; the historian client is unaware data was tiered.
Query & bill
Analytic and threshold queries answer from the model. Billed per TB-compressed.
Pilot it on your historian
Bring a real export from PI, AVEVA, Insights Hub or an OPC-UA tag stream. We’ll report the ratio and the byte-exact result on your own data, on-prem.
Patent pending (US provisional filed 2026). On-prem encoder is a proprietary compiled binary (license-gated via the engines dashboard); free, open decoder (Apache-2.0).