ML pipelines
Preprocess on your machines. Ship vectors, not data.
Preprocess images, video, documents, and CSV rows on your infrastructure. Only embedding vectors ever leave the pipeline — your raw data stays put, and every run is signed and reproducible.
Raw media never leaves the machine. Only vectors move.
Preprocessing that
stays in place
Image, video, document, and row preprocessing on your own infrastructure. The pipeline output — vectors — is all that moves.
Signed & Reproducible
Plugins are Ed25519-signed and versioned, with full lineage. Re-run any pipeline and get the exact same vectors.
In-Place Preprocessing
Raw images, video, and documents stay on your storage. Load, resize, and embed execute where the data lives.
Sandboxed Execution
Every plugin runs in an isolated sandbox with seccomp. The pipeline is deterministic and auditable, start to finish.
How it works
in production
A pipeline that turns raw media into vectors on your machines, with nothing extracted.
Raw data arrives
Images, video, documents, or rows land in the watched directory on your storage.
Agent detects the workload
The agent reads the file signatures and picks the matching pipeline.
Loads the media
The load plugin reads the raw bytes in place — no copies, no staging.
Preprocesses locally
Resize, crop, and normalize run on your hardware, in the sandbox.
Embeds and exports vectors only
The embed plugin produces 809-dim vectors, written to your output store. Raw media never leaves the machine — untouched, in place.
Validated where
it runs
Proven on production media workloads — with every step logged and auditable.
Only embeddings leave the pipeline. Raw media stays in place.
Signed, versioned plugins with full lineage. Same input, same vectors.
One self-hosted agent with a 12–15 MB footprint.
One immutable record of every preprocessing step and result.
Plugins this solution
needs
Signed, sandboxed, and versioned. SentraZero runs them where your data lives — every decision logged and auditable.
loadReads raw media and documents in place. No copies, no staging.
inraw media → prepared bytes
{ "batch": 64, "in_place": true }
✓ signed✓ sandboxedsha256:d848f444…
resizeResizes, crops, and normalizes media to the pipeline spec.
inprepared bytes → spec crops
{ "size": "512x512", "crop": "center" }
✓ signed✓ sandboxedsha256:c3c1453f…
embedProduces 809-dim vectors from the prepared media. Raw files never leave the machine.
incrops → 809-dim vectors
{ "dim": 809, "model": "sentra-v1" }
✓ signed✓ sandboxedsha256:8482d73b…
Explore other
solutions
Every solution runs on the same self-hosted engine. Your data stays where it lives, no matter the workload.
Preprocess where
your data lives
Show us a sample dataset. We'll stand up an agent on your infrastructure, run a real preprocessing pipeline on your data, and hand you the vectors plus the audit trail. Nothing leaves your storage.