Pipeline metrics
Real benchmarks, zero client data
Three production pipelines, measured where they run: retail catalog comparison, image embedding, and duplicate detection. Every number below came from a real run on client infrastructure — with zero client bytes in the control plane.
Measured in
production
Three pipelines, three workloads, one guarantee: your data never leaves your storage. Here's what they look like when they run.
Retail Catalog Compare
Scrapes 9 URL pairs and compares 41 SKUs in about 65 seconds total — with a verdict and a reason for every row.
Image Embedding
Preprocesses 202 images into 809-dim vectors in 27–81 seconds per batch. Raw files never leave the machine.
Duplicate Detection
Reconciles 19 products and flags hidden duplicates in 14.4 seconds. Zero rows of client data leave the pipeline.
How it works
in production
Every measured run follows the same audited path — on the machine where the data lives.
Job created
A pipeline is created from the solution you picked — retail catalog, image embedding, or duplicate detection.
Plan is claimed
The control plane publishes the job; an agent on your infrastructure picks it up.
Sandbox starts
The agent applies seccomp and verifies the plugin signatures before anything runs.
Workload executes
Scraping, embedding, and reconciliation run entirely on your hardware.
Result written, audit logged
Outputs land where the data lives — raw files never leave the machine. Every run appends to the immutable audit trail: plugin, version, duration, checksum.
Validated where
it runs
Real numbers from real production runs — not synthetic benchmarks.
9 URL pairs scraped, 41 SKUs compared, verdict + reason per row.
202 images → 809-dim vectors per batch. Raw files stay put.
19 products reconciled, hidden duplicates flagged.
Zero client bytes in the control plane — by construction.
Plugins this solution
needs
The same signed, sandboxed plugins behind every measured run.
scrapeFetches live product pages for comparison. Automatic retry keeps the run green.
inurls → product snapshots
{ "retries": 3, "timeout_s": 15 }
✓ signed✓ sandboxedsha256:10605d28…
compareMatches vendor and marketplace listings with similarity thresholds and a UPC fast-path.
insnapshots → verdicts + reasons
{ "threshold": 0.82, "upc_fastpath": true }
✓ signed✓ sandboxedsha256:ffbba70d…
loadReads raw media in place for embedding. 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 prepared media. Raw files never leave the machine.
incrops → 809-dim vectors
{ "dim": 809, "model": "sentra-v1" }
✓ signed✓ sandboxedsha256:8482d73b…
dedupFlags the same product hiding under different IDs before it doubles ad spend.
incatalog rows → unique rows
{ "keys": ["brand", "model", "sku"] }
✓ signed✓ sandboxedsha256:6c652181…
Explore other
solutions
Every solution runs on the same self-hosted engine. Your data stays where it lives, no matter the workload.
See it on
your workload
Bring a real workload. We'll stand up an agent on your infrastructure, run one of these pipelines on your data, and show you the timings plus the audit trail. Nothing leaves your storage.