Available now
InertialAI-Embed-Alpha
A multi-modal embedding model for signals. Encode each input — a time-series on its own, or a time-series paired with a text description — into one shared space, then cluster, search, classify, and retrieve.
Embed and Cluster Signals
Built on Chronicle and finetuned with contrastive learning, Embed-Alpha places signals with similar shape and meaning next to each other. Pairing each series with a short text description — a multi-modal input — sharpens those clusters further, lifting 1-NN retrieval accuracy by 2–5 points on standard benchmarks.
ItalyPowerDemand (0.948) and Wafer (0.991) are already near ceiling on series alone; multi-modal helps most where the text adds context the signal lacks.
Cross-Modal Retrieval
Because text and time-series share one space, you can also retrieve across modalities — query a corpus of signals with natural language, or label a signal by its nearest text. It is one capability of the model, not the whole story.
Matryoshka Embeddings
Trade dimensionality for efficiency without retraining. Smaller embeddings keep most of the clustering quality, so you can shrink storage and speed up nearest-neighbor search.
Use Cases
Cohort discovery
Cluster patient monitoring traces — with or without clinical notes — to surface groups that share a condition without hand-labeling.
Fleet & sensor grouping
Group machines by vibration or temperature signature, pairing each series with its maintenance report to tighten the clusters.
Signal classification
Nearest-neighbor classify new signals against a labeled bank — 88–99% 1-NN accuracy across standard UCR/UEA datasets.
Unified observability
Cluster logs and metrics together, or query them with natural language through the shared cross-modal space.