A national satellite pay-TV provider asked us to replace manual set-top box testing with automation they could trust. In twelve months we designed, built and deployed an air-gapped, GPU-accelerated verification platform in their device lab. It presses real remote buttons, watches the real video output, judges what it sees and hands engineers an evidenced verdict. By the end of the year the whole QA team could test at once, the client was building its own vision models in-house, and every result could be read in plain English.
| Client | National satellite pay-TV provider (name withheld) |
| Problem | Manual, unrepeatable certification of set-top box firmware across six product lines |
| Deployment | Air-gapped, on-premise NVIDIA GPU fabric inside the client's device lab |
| Duration | 12 months, from first design to lab-wide operation |
| Design scale | 128 boxes per rack · 50 racks · 6,400 devices · 300 ms from frame to verdict (design targets) |
The lab certifies every firmware build for six product lines, each with several versions in test at once. The goal was never to review faster. It was to make exhaustive certification cheaper than sampling.
Coverage was whatever one person could get through in a day, and no two runs were quite the same.
Earlier tools compared screenshots pixel by pixel. A font tweak or a new promo tile broke the test, so most failures were the tool's fault.
A note in a spreadsheet is a recollection. An annotated proof frame and a log start a different conversation with a developer.
One capture stream showed sixteen boxes, but the audio path reached one box at a time, so audio wasn't verified at scale at all.
| Design target | Per stream | Per rack (128 boxes) | Full build-out (50 racks) |
|---|---|---|---|
| Devices under test | 16 | 128 | 6,400 |
| 4K capture streams | 1 | 8 | 400 |
| Device-lane frames per second | 960 | 7,680 | 384,000 |
| Uncompressed video | ~6 Gbps | ~48 Gbps | ~2.4 Tbps |
| Capture → verdict | 300 ms | 300 ms | 300 ms |
Engineering estimates at 60 fps, 8-bit 4:2:0 video · design targets, not measured
At terabits per second of decoded video, frames can't be copied through the CPU. Every step has to run in GPU memory.
Freeze, macroblocking, stuck splash screens and audio dropouts only show across a sequence of frames.
Six product lines × several firmware versions: the same check has a different correct answer on each.
Every verdict starts with a real IR keypress. Thousands of remotes need scheduling, with hardware locked so jobs never collide.
Drives the remote, reads the screen with detection and OCR, judges A/V health on the GPU and attaches proof frames and a written report to every verdict. Sixteen boxes per 4K stream, sliced on the GPU. Tests are configuration, not code.
An on-premise path from raw footage to labelled dataset to trained, versioned model. The client now retrains its own detectors: mAP rose from 0.836 to 0.961 in one cycle.
Channels are matched against a managed library of reference images, so a new or rebranded channel needs a picture, not a retrain. The library rebuilds in under two seconds. Below a confidence margin the answer is “unknown” and a person looks.
The whole QA team tests at once across 8 equipment rooms, with hardware locked per job, per-user attribution, a full audit trail and inventory synced from the client's own configuration database.
Plain-English reports and plain-English test authoring, on separate inference hardware inside the lab. Facts such as channel numbers come from configuration, never from the model.
Networked control of the audio routing lets one scenario step cycle through every box and check measured signal level, so audio is certified on all sixteen paths.
| Outcome | What changed |
|---|---|
| Automation the team trusts | Interpretation replaced screenshot comparison, so harmless UI changes no longer cause false failures. |
| Evidence on every verdict | Pass, fail or partial, each with an annotated proof frame and a per-lane log, one click away. |
| In-house model ownership | The client builds, versions and retrains its own models. |
| Flat cost as the line-up grows | New and rebranded channels take minutes, not a retraining project. |
| Fails safely | Low-confidence frames return “unknown” and go to a person. The system never guesses. |
| Data sovereignty | Fully air-gapped. No footage, labels, models or results leave the lab. |
This project was in media device testing, not manufacturing. What carries over is how it was engineered.
GPU video pipelines that judge what they see, not pixel diffs, are the core of our Vision Quality Inspection kit and DeepVision at the cell.
Every verdict carries its proof. In AgentMesh that becomes the signed Evidence Log and the nightly evidence pack.
Where the business changes often, the change goes into managed reference data, not model weights. The same idea runs through Vector Memory and the frozen-base doctrine.
Nothing left the lab. Nothing leaves your site either.
Client details are withheld under confidentiality. Year-one results were measured in the client's lab; scale figures are engineering design targets.
