The mismatch
Machines want everything.
Each era of data systems was built around a different consumer. The third and current phase is the first where the consumer isn't human.
One project tree
Every modality.
One query language.
Video, images, audio, embeddings, point clouds, sensor streams, tables, text — addressable in one tree, queryable in one language, governed by one permissions model.
Object-store native. Hybrid by default — one deployment spans your neocloud, your tradcloud, and your on-prem cluster.
Declarative samples
A sample is a query.
Spiral stores immutable raw bytes on typed coordinate frames. Your training sample — a window, a resample, a transform — is declared, not denormalized, and materialized on demand.
Reproducibility is free. asof(T) rewinds the whole project — every modality, every alignment — to any point in its history.
Coordinate-aligned streamsT
Video30 Hz
Audio48 kHz
IMU200 Hz
LIDAR10 Hz
sample(window=200ms, align=“nearest”)
Field Guides
How Spiral handles visual, spatial, and embodied data.
Follow real specimens from source formats and coordinate systems to the physical reads and model inputs Spiral produces.
Working with video data
Solution
Everything but data acquisition.
Frames, clips, samples, embeddings, and every access pattern in between. Your videos, training-ready.
Deep dive
Machines do not press play.
An interactive teardown of how Spiral reads MP4, from the codec graph and byte ranges, through hardware decode, to model-ready tensors. One modality, end to end.
Saturate the accelerator
Compressed bytes,
model-ready tensors.
Object storage to model-ready tensors, as one pipeline — every modality, straight into accelerator memory. The format underneath is Vortex: the open columnar format we created and donated to the Linux Foundation.
Object store
GPU tensor
Object storage stays the source of truth. Spiral materializes the shape your workload actually asks for — random reads for retrieval, aligned windows for training, full scans for analytics.