One company.
Two independent engines.
Vectorio builds two products. Protor moves and computes your data; Indago finds what you've already read. Same sovereignty principles, different problems — take either, or both.
Vectorio Protor
The deterministic data engine. Streaming ETL, distributed SQL, columnar kernels, plugin architecture. Built for the work that must be reproduced under examination — at the speed your hardware can actually deliver.
Inside Protor → Research · KnowledgeIndago
A private research engine for the question: how do you find the one filing you forgot you'd read? Indexes filings, papers, code, correspondence — locally — into one ranked, citation-grade search. A different product. A different problem.
Open Indago ↗What others need a cluster and a cloud for.
We do on one machine.
Your most valuable data can't go to the cloud — but the best tools require it. Protor closes that gap: cluster-class analytics on hardware you own, with proof for every result. It's the quadrant no incumbent occupies — and the reasons compound.
Cluster-class on one machine — even a Raspberry Pi
105–118 GB/s on a single Apple M3 Max; the same engine runs on a $-class edge device. No cluster, no per-query bill — the infrastructure others rent by the hour, you own.
Provable by architecture — not bolted on
Bit-exact deterministic replay (SHA-256), per-row access control, cryptographic audit — built into the engine. Spark/Databricks are non-deterministic by design and can't retrofit it.
Your data never leaves
Offline-first, air-gap-capable, EU-built. Sovereign by architecture, not by a contract clause — the right side of every data-residency rule.
The closed loop no one matches
Observe row data inline → decide → act on the engine (pause, isolate, quarantine) within the batch, sub-millisecond, autonomous, fully audited. Competitors can observe or act — not both, inline.
Protor — touch your hardware ceiling.
A 1-billion-row analytical matrix on a single Apple M3 Max — grouped and summed in 32 ms, on one machine. The overnight report finishes before standup: no cluster, no rented compute. Absolutes, not ratios — and we publish the caveats with them.
Short-key int32 SUM over 1 B rows · 32 ms · compressed-int kernel, no spill (g11).
town × SUM(price) + COUNT over UK property data — 25× faster than the prior baseline.
SUM + MIN + MAX + COUNT in a single scatter pass (gs06). Real engineering ceiling, not a synthetic best.
One .NET process. No GPU. No cluster. NEON SIMD on Apple Silicon. Same binary runs on Linux x86-64.
Apple M3 Max · 16 cores (12 P + 4 E) · 64 GB unified memory · single .NET process · NEON SIMD.
Sample projects, real pipelines.
Every project below runs on Protor, the Vectorio data engine. We open-source the workflows so you can see exactly what they do — node by node, byte by byte.
Beethoven, jitter-free.
A real-time audio pipeline that streams Beethoven through up to 10,000 audio nodes — EQs, compressors, reverbs, delays, gates, limiters — without jitter or a single dropped sample. Determinism at the speed of sound.
Munich, cooled.
A Sentinel pipeline mapping urban heat across Munich: ingest Sentinel-2/3, GDAL reproject, mask cloud, compute thermal indices, aggregate per neighbourhood, ship a signed GeoJSON map.
Mailbox, sealed.
A compliance-grade mail-archive pipeline: ingest IMAP, redact PII, hash every message, ship to a signed evidence pack an auditor can replay.
Tissue, classified.
Camelyon16 ONNX whole-slide pipeline: ingest pyramidal TIFFs, classify against ONNX tumour models, ship signed CSV per slide.
Findings, ledgered.
VSNEP: a NIS 2 evidence pipeline for essential entities. Ingest Suricata IDS alerts + ASA vulnerability findings, normalise, hash-chain into a tamper-evident ledger, ship a self-verifying bundle.
Weather, windowed.
A live weather pipeline: subscribe to a REST API, paginate paginated history, aggregate over a rolling window, push to downstream consumers. Same orchestration pattern any time-series feed needs.
Faces, detected.
A real-time webcam pipeline: ingest video, detect faces, overlay landmarks, optional live face-blur for privacy. Same engine drives RTSP surveillance and forensic video evidence.
Prompts, local.
An on-device LLM pipeline: stream CSV prompts, run them through TinyLlama on the host CPU, output answers to CLI. No cloud. No API key. Never leaves the box.
Want to see one work?
A 30-minute call. We'll show you a real pipeline running on your data shape — not a sandbox demo. On your benchmarks, on commodity hardware. We come back with the measurements.