One encoder.
Storage, analytics and governance together.
Polygen fits a small model to each window of your data and stores it as a token. The token is smaller, and it stays searchable, auditable and ready to train on. Raw records never leave your environment.
Problem
Data infrastructure wasn't built for the volumes you're running today.
Storage, transmission, and processing layers were designed independently. Each one solves its own problem. As volumes compound, the cracks between them turn into cost. You pay to store the data, you pay again to move it, and you pay again to prepare it before anything useful runs on it.
Conventional compression helps with the first line and breaks the next two. Compressed bytes are opaque, so analytics and AI force a full decode before any work can run. The compression and the work cancel each other out.
Conventional compression helps with the first line and breaks the next two. Compressed bytes are opaque, so analytics and AI force a full decode before any work can run. The compression and the work cancel each other out.
Solution
One encoding layer. Three immediate wins.
Datasent encodes any data, tabular, sensor, time-series, images, and video, into a single lossless token format. The same representation delivers across storage, transmission, and compute.
Storage
Compress Losslessly
Lossless tokens shrink tabular and telemetry data 5 to 11x, measured, with every byte exactly recoverable. Coefficient-only archives reach 50 to 250x on sensor streams, and up to 3,000x on smooth sensor data.
Lossless tokens shrink tabular and telemetry data 5 to 11x, measured, with every byte exactly recoverable. Coefficient-only archives reach 50 to 250x on sensor streams, and up to 3,000x on smooth sensor data.
Bandwidth
Raw data stays local
A shared model basis is agreed once between environments. After that, only the residual (the unpredictable part) crosses the network. The original is reconstructed exactly on the other side. Raw data never leaves the source environment.
A shared model basis is agreed once between environments. After that, only the residual (the unpredictable part) crosses the network. The original is reconstructed exactly on the other side. Raw data never leaves the source environment.
Analytics & AI
Skip the decode step
Tokens are sufficient statistics for trend analysis, anomaly detection, similarity search, model training, and inference. Models approach raw-data accuracy at a fraction of the data volume, latency, and cost.
Tokens are sufficient statistics for trend analysis, anomaly detection, similarity search, model training, and inference. Models approach raw-data accuracy at a fraction of the data volume, latency, and cost.
Proven on vision-language models and multimodal RAG.
The same encoder compresses the visual tokens a vision-language model attends over. Measured end to end on Qwen2-VL-7B, on public benchmarks, against the un-spliced baseline.
+8.4 points of accuracy
ScienceQA on Qwen2-VL-7B at the recommended operating point, with 41% fewer attention FLOPs. One rank-16 LoRA trained on 300 images, and it transfers to MMMU, VQAv2 and DocVQA without retraining.
~71% fewer attention FLOPs
At the compression operating point accuracy is still +7.4 points and the stored visual representation is 7.5x smaller. A FLOP count, not wall-clock: single-shot wall-clock savings are not claimed.
Encode once, query many
A visual corpus stored as polygen descriptors is 5.5x smaller than cached fp16 embeddings (10.4 MB against 56.7 MB), answers the same questions, and serves about 3.7x more requests per GPU on video frames.
Process
How Datasent works
Lossless
Every byte recoverable.
Data is segmented and fitted against a deterministic basis. The unpredictable part (the residual) is stored exactly. Reconstruction is mathematically exact, not approximate. If the basis captures nothing, the residual still recovers the original at full fidelity.
Local
Raw data stays put.
Both environments agree on the basis upfront. Only the residual and a small metadata payload travel. The basis carries most of the information and is regenerated on each side rather than transmitted. Raw records never cross a network boundary.
Limitless
Usable, not just stored.
The coefficient matrix inside each token is a sufficient statistic for trend analysis, anomaly detection, similarity search, model training, and inference. Analytics and AI run on the tokens directly, with no decode step in front of every workload.
Use Cases
Who it's Built For
Business & Enterprise
Understand data opportunities that were previously out of reach.
Explore product
White Paper
Deep dive into Datasent's approach
The white paper covers the full mathematical framework: lossless tokenization, adaptive
segmentation, token-space computation, and zero-knowledge proof compatibility. No proprietary
dependencies. No lossy trade-offs.
segmentation, token-space computation, and zero-knowledge proof compatibility. No proprietary
dependencies. No lossy trade-offs.
The next number should come from your data.
Every number on this site is a measurement with its scope attached. We would rather measure your data than estimate it: a 30-minute scoping call maps your data shapes and constraints to a concrete pilot, and a hosted evaluation instance follows in days, with nothing to install.
Questions? Reach us at info@datasent.com
Join our newsletter
Get updates on research, product news, and privacy-first data insights.
