Hidden spatial redundancy
Sensors capture thousands of near-identical frames that bloat datasets without adding a single bit of new information.
80%+ of real-world data
One integrity engine — Medical imaging. Spatial perception. Digital content.
An asynchronous, high-throughput perceptual data integrity engine. Built for 3D DICOM volumes, high-frame-rate spatial sensor streams, and digital content pipelines. Eliminate data leakage and purge redundant data before it ever reaches storage or a training bucket.
Engine domain
Status: ingesting…
{ "session_id": "vxl_89234", "engine": "Health_3D", "frames_processed": 1420, "redundancy_detected": 0.42, "status": "DEDUPLICATED", "integrity_score": 0.998, "savings": { "storage_cut": "34.2%", "annual_usd": "$12,402" } }
The petabyte bottleneck
A unified data integrity platform for high-stakes AI
Modern AI depends on massive volumes of imaging, sensor, and spatial data — yet more than 80% of real-world data is redundant, duplicated, or structurally identical. Voxelion's fingerprinting engines detect and remove those duplicates across 2D, 3D, and spatial sensor streams, so models train only on unique, high-value scenarios.
Model bias & false accuracy
Duplicates split across train and validation leak answers. Models memorise specific voxels instead of learning generalised features.
Silent validation leakage
Millions burned on cloud & GPU
Storage and training cost scale linearly with raw volume, yet a large share of that data provides zero marginal information gain.
Cost scales with noise
Purpose-built domains
Three engines. One platform. Zero data leakage.
AI-native data integrity for healthcare, autonomous perception & digital content.
Voxelion Health
Volumetric medical data integrity for clinical AI
Generates a 256-bit perceptual fingerprint for every sub-volume in a 3D DICOM dataset, computed from voxel data alone. Near-duplicate frames are identified and filtered before they reach a labelling queue or a training split.
- 30%+ reduction in storage
- 10× faster data preparation
- FDA-ready audit trails
- Medical foundation model training
Voxelion Drive
Spatial sensor deduplication for autonomous driving & robotics
Applies the same perceptual fingerprinting core as Voxelion Health to LiDAR point clouds, camera frames, and radar sweeps — across all sensor modalities in a single index. Redundant captures are flagged and removed before they reach cold storage or a perception model training run.
- 30%+ reduction in GPU costs
- Faster model convergence
- Synthetic scenario curation
Voxelion Curate
Perceptual integrity for digital content pipelines
Applies the same perceptual fingerprinting core as Voxelion Health to product image archives and generative AI training corpora. Near-duplicate content is flagged and filtered before it reaches a marketplace listing or a model training bucket.
- Marketplace anti-counterfeiting at the image layer
- GenAI corpus deduplication — 40–70% typical redundancy
- Auditable dedup trail for copyright defensibility
Architecture
The data integrity layer AI has been missing.
AI models fail not from a lack of data, but from too much redundant data. Voxelion acts as a pre-storage filter, so only unique, high-entropy data reaches your buckets.
Perceptual fingerprinting
A compact signature per image, derived from its visual structure rather than its bytes. It widens automatically for volumetric DICOM, where the standard width is measurably too narrow to keep distinct patients apart.
Sub-microsecond comparison
0.64 µs per comparison, measured — about 1.6 million per second per core. Large corpora are indexed, so a lookup touches a candidate set rather than everything.
Identity from pixels, not headers
Identity comes from pixel structure, not from headers that can be rewritten or stripped. A renamed, recompressed, re-encoded copy still matches.
import os, requests
# Plain HTTP and a bearer token. No SDK required.
files = [("files", (p, open(p, "rb"))) for p in paths]
r = requests.post(
"https://api.voxelion.ai/api/dedup",
headers={"Authorization": f"Bearer {os.environ['VOXELION_API_KEY']}"},
files=files,
data={"engine": "curate", "maxDistance": 10},
)
r.raise_for_status()
report = r.json()
# Voxelion reports. It never deletes — acting on the
# keep/drop manifest stays your decision.
print(report["leakage"], len(report["clusters"]))Official Python and TypeScript SDKs are on the way — see the plan.
-
0.64 µs
Per comparison, measured
-
1.6M/s
Comparisons per core
-
0 bytes
Uploaded in local mode
-
100 MB
Free in your browser, no account
Usage-based · From free
Three engines. Three pricing models.
Priced in the unit native to your domain — per scan, per vehicle or per check. Free evaluation tiers are available for all three engines with no credit card required.
Voxelion Health
$199/month
Monthly subscription · 5,000 scans/month
- Priced per scan processed
- Free evaluation: 500 scans/month
- Academic free: 10,000 scans/month
Voxelion Drive
$750/month
AV fleet subscription · up to 5 vehicles
- Priced per vehicle, project or volume
- Geospatial projects from $100
- Free evaluation: 500GB lifetime
Free KYC pilot
Voxelion Curate
$150/month
Marketplace subscription · 5,000 comparisons/month
- GenAI corpus audits from $500 flat
- Banking and KYC from $0.05/check
- Free pilot: 1,000 checks for 30 days
Usage-based pricing from $0 · Three engines · Three pricing models. All pricing is exclusive of applicable taxes.
Not sure where to start? Fixed-fee, 30-day enterprise pilots are available across all three engines for $5,000–$15,000 depending on engine and dataset volume. Contact partnerships@voxelion.ai to scope your pilot.
Engineering FAQ
Questions perception teams ask first
How is this different from file hashing or checksums?
Checksums only catch byte-identical files. Voxelion fingerprints the spatial content itself, so it still detects near-duplicate frames after re-encoding, metadata rewrites, or minor sensor noise.
Does Voxelion need access to patient data or PHI?
Fingerprints are computed from voxel intensity alone, so no patient identifier ever enters the match. DICOM headers are read only to group slices into series and studies, and anonymised or pseudonymised identifiers work for that. If headers are stripped entirely, deduplication still runs, but slices are no longer grouped into volumes.
What throughput can the engine sustain?
A single comparison takes about 0.64 microseconds — roughly 1.6 million per second on one core, measured rather than estimated. Comparison is not the bottleneck in practice: decoding is, and large corpora are indexed so that a lookup touches a small candidate set instead of the whole library. Large batches can be queued asynchronously, so deduplication never sits in the path of an ingest.
Can I run Voxelion inside my own VPC?
Yes. The Enterprise plan supports on-premise and private-cloud deployment, with custom engine tuning and dedicated SRE support.
Which formats are supported?
Voxelion Health handles 3D DICOM and NIfTI volumetric imaging. Voxelion Drive handles LiDAR point clouds and high-rate camera or radar frame streams. Voxelion Curate handles 2D raster content — JPEG, PNG, WebP and TIFF — for product image pipelines and GenAI corpus curation.
Is there an SDK or a CLI?
Not yet. The API is plain HTTP with bearer authentication and JSON responses, so any HTTP client works today — the documentation includes a complete working example in about fifteen lines of Python. Official Python and TypeScript SDKs and a downloadable CLI are planned; the plan is published rather than announced as available.
Does Voxelion delete my data?
No. Voxelion reports what is redundant and hands back a keep/drop manifest; removing anything is your decision and happens on your side. Uploads are deleted after a run completes — by settlement time there is nothing left on our disk. In the console's local mode and the no-account demo, files never leave your browser at all.
What content types does Voxelion Curate support?
Curate currently handles 2D raster formats: JPEG, PNG, WebP, and TIFF. It operates on individual images or batched corpus directories via the REST API or Python SDK. Video frame extraction pipelines are on the roadmap for H1 2027.
How does Curate handle generative AI training data without accessing proprietary model weights?
Curate operates entirely at the data layer — it never touches model weights, training code, or infrastructure. It receives a stream of image files or a directory manifest, generates a perceptual fingerprint per file, compares against the corpus index, and returns a deduplication report. No model access required.
Deploy in minutes
Ready to eliminate data leakage and accelerate AI performance?
Request a demo of Voxelion Health, Voxelion Drive, or Voxelion Curate.
Fingerprints from voxel data · No PHI required