Deficit-weighted allocation
Each path is scored by goodput × urgency × cost. Work goes to whichever path will run dry soonest. Cost is a mild tiebreak only; never a reason to pick a cheap slow path over a fast one mid-transfer.
MVRK Systems builds AI deep-tech infrastructure. In practice that means software that models every route it can reach (wired, cellular, satellite) and decides, without being asked, where each chunk of a shoot should go. Then keeps deciding as cables are unplugged, towers congest, dishes obstruct, and the deadline moves closer.
Bonding a 400 Mbps tower to a 20 Mbps dish is not a matter of splitting the file in half. Most of the work is in prediction and failure handling; estimating what each link will do next, and not mistaking a slow path for a broken one. No model is trained on your footage; the intelligence is in the decision-making, and it runs entirely on the box.
Each path is scored by goodput × urgency × cost. Work goes to whichever path will run dry soonest. Cost is a mild tiebreak only; never a reason to pick a cheap slow path over a fast one mid-transfer.
The stall threshold scales with the work in flight. A 64 MB chunk on a 20 Mbps dish legitimately takes 26 seconds, so a fixed timeout would bench a perfectly healthy satellite link on every single chunk.
A stalled path revives after twenty seconds and re-enters with small trial chunks. Without it, one tunnel or one underpass removes a path for the rest of the shoot.
Past 95% complete, the oldest outstanding chunk is speculatively re-issued to any healthy path at least 1.5× faster. It costs a few hundred megabytes and kills the "99% done, waiting on one straggler" failure.
Chunks are tagged ANY, CHEAP_ONLY, or FAST_ONLY, and routes are filtered by affinity before scoring. This is what guarantees independent mode cannot quietly spend cellular data.
A usable path idle for more than five seconds is force-fed work so its estimate cannot go stale, and no single throughput sample may exceed twice a path's observed peak; otherwise one lucky sample starves every other link.
AGGREGATION PERFORMANCE IS MEASURED ON HARDWARE, NOT SIMULATED · FIELD TEST PENDING
A single connection is a single point of failure. The edge node treats every available link as one lane of a wider road (a house Ethernet line, two cellular carriers, a satellite dish) and the cloud node is the only place the file exists whole again. Whatever the location gives you, the box uses it.
The card mounts over a reinforced USB-C host port and stages to the internal NVMe cache as fast as the card will read; the card sets the ceiling, not the appliance. Staging is local and complete before a single byte leaves the box, so a dropped link never costs the operator a re-copy.
Files are cut into encrypted 64 MB chunks and queued per route. Each chunk carries its own ordinal and hash, which is what allows four links of wildly different latency (2.5 GbE wired down to a contended dish) to be treated as one.
Wired Ethernet takes the bulk whenever a house line exists; two bonded 5G modems carry it when one doesn't; satellite covers the rest. Throughput on each lane is measured continuously and the queue re-weights against it; a congested tower or an unplugged cable degrades the run, it does not stop it.
The regional cloud node reassembles chunks in ordinal order and checks a SHA-256 against the source manifest. Only a verified file triggers the delivery webhook into Frame.io, Air.ink, or an S3-compatible target. An xxHash64 sidecar ships alongside so the receiving house can check the media with the tools it already runs.
The policy engine re-evaluates every five seconds against queue composition, available routes, deadline, and the cellular budget, then re-tags pending chunks. A thirty-second minimum hold stops it flapping between modes on a marginal link. Plugging in a network cable mid-shoot is a configuration change everywhere else; here it is just a route appearing.
All routes serve one queue.
Chosen when proxies are outstanding, the deadline is at risk, or there is no flat-rate route available. Maximum speed, maximum cellular spend.
RAW pinned to the flat-rate route.
Chosen when proxies are clear and a wired or satellite route is healthy, with no deadline pressure. Cellular carries only priority traffic, so the data bill stays flat; and on a house Ethernet line it stays at zero.
Metered route idle until the others fail.
A dashboard override for a borrowed dish or a capped line. That route stays parked and only wakes if the routes carrying the queue drop.
The product promise is that the editor starts cutting within minutes of the card going in. Priority classes are what make that survivable: the proxy queue drains first regardless of how much RAW is stacked behind it, and affinity rules mean no amount of spare cellular speed can pull a chunk off the path it was pinned to.
Media that arrives corrupt is worse than media that has not arrived, because someone will cut with it. Verification is not a feature here; it is the gate.
Colour on an MVRK device is a language, reserved for state and never decorative; and so is behaviour: solid, pulsing, or blinking each mean something. A DIT in the rain at 2 a.m. should be able to read the run from six feet away without taking a glove off.
A separate indicator shows how many routes are live: wired, two cellular, satellite. A dot is bright when it is actively carrying data and dim when the route is present but parked, which is how independent and standby modes read from across a set.
Anything with a menu invites configuration, and configuration on a wet camera cart at 2 a.m. is how footage gets lost. The plate reports state; the producer sets the deadline and the budget once, in the dashboard, before the shoot starts.
DailiesDrop™ proves the orchestration layer. Everything after it reuses the same transport, the same verification, and the same status language on a different job.
Zero-UI edge ingestion. Card in, cloud out, at card speed across a bonded uplink.
Specification → Beta 0.1.0Naming and delivery conformance. The convention written down, checked against the media, and sent with the drive so the far end can check it too.
What it checks →Automated live-show showcalling. Cue stacks driven from the same orchestration layer that moves the media, so the show and the media never drift apart.
Real-time sports telemetry. Field-side capture and distribution for broadcast rights-holders, on the same chassis discipline.
Dead cell zone, terabyte days, a delivery deadline in another time zone. Those are the interesting ones.