How plant quality engineers can run factory visual inspection

How can the inspection decision stay at the line?

A cloud dashboard should not be the only place a line can decide what to do with a part. Define what happens when a camera misses a trigger, inference takes too long or the network is unavailable. For a first pilot, run the system in observation mode and let qualified operators make production decisions.

The diagram describes a proposed system, not a validated HyperLake product bundle.

The edge application handles capture, preprocessing and inference. The cloud path manages selected evidence, review records and model lifecycle. Model releases travel back through a controlled change process. A cloud outage must not silently change a line's acceptance rules.

Which components need evaluation?

Basler lists 1920 × 1200 resolution and a 42 fps frame rate for this camera. Those are camera specifications, not achieved inspection throughput. NVIDIA describes DeepStream as a streaming analytics toolkit; its availability does not establish that a particular model meets this line's deadline.

What storage might the evidence require?

Worked estimate, not a benchmark: assume four cameras, 1920 × 1200 pixels, an RGB8 working representation of three bytes per pixel, 10 captured frames per second per camera and 16 operating hours per day. The resolution comes from the camera specification; camera count, representation, rate and operating hours are scenario assumptions.

Uncompressed frame bytes are 1920 × 1200 × 3 = 6,912,000. The four cameras produce 4 × 10 × 16 × 3,600 = 2,304,000 frames per day. Retaining every RGB working frame would therefore consume 15,925.248 GB per day, using decimal GB and excluding headers, metadata, replicas and temporary files.

If the workflow retains only 1% of those uncompressed frames, the image subtotal is 159.25248 GB per day. That percentage is a retention-policy assumption, not an expected defect rate or compression result. Keep a representative random sample as well as flagged images; otherwise future evaluation can miss systematic false negatives. Measure the encoded bytes on actual images before sizing the final store.

How should the line timing budget be set?

A proposed timing budget can be derived rather than advertised. If a downstream actuator is 0.4 metres from the camera and the conveyor moves at 1 metre per second, travel time is 400 milliseconds. Reserving 150 milliseconds for actuation and margin leaves a proposed 250-millisecond deadline for capture, processing and dispatch. These are assumed geometry and budget values. Measure end-to-end latency and missed deadlines; neither cloud latency nor model-only speed proves compliance with the budget.

What is a sample compute cost per camera?

Illustrative compute-module estimate, not a complete inference cost or quote: assume one NVIDIA Jetson Orin NX 8 GB module per camera, amortized over three years, a 40 W module power mode, 16 operating hours per day and 250 operating days per year. NVIDIA lists a $649 USD volume suggested price at 1,000 units and describes the 40 W mode for Orin NX production modules; these inputs do not include a carrier board or a complete industrial computer. The U.S. Energy Information Administration reports a 2025 industrial electricity average of 8.51 cents per kWh; a local tariff can differ.

Module allocation = $649 ÷ 3 = $216.33 per camera-year. Assumed module electricity = 0.040 kW × 16 hours/day × 250 days/year = 160 kWh per year, or 160 × $0.0851 = $13.62 per camera-year. The subtotal is $229.95 per camera-year. It excludes camera, lens, lighting, carrier board, enclosure, power conversion, other system load, installation, cooling, cloud, storage, software, support and tax. It does not show that the module can meet the assumed camera workload; measure the complete design before sizing or procurement.

For an annual site estimate, use camera count × (a customer quote for each complete edge computer divided by its chosen service life + measured annual computer kWh × the site's tariff), then annualize quoted one-time cameras, optics, mounts and installation over their chosen service lives; add annual recurring labor, software/support, cloud/storage/egress and tax as applicable. Use the module-only floor above only when no complete-computer quote exists; do not add it again if the quoted computer includes that module. No cost per inference is claimed without measured throughput.

What safety, privacy and security responsibilities remain?

Keep quality inspection separate from personnel protection. In the US, OSHA's machinery-guarding requirements address protecting operators and other employees. This proposed vision pipeline is not represented as a certified guard or safety control. For Canada, have the responsible site team assess applicable provincial or federal workplace requirements.

Frame cameras around the product rather than workers. Where employee information is captured, review applicable privacy law, purpose, notice, access and retention. Canada's privacy commissioner has specifically addressed employee notice for employers subject to PIPEDA; the rule's scope should not be generalized to every employer. Review US state and employment requirements with the appropriate adviser.

Use the NIST operational-technology security guide to inform the security review. Separate inspection services from production controls and assign owners for remote access, patching and recovery. No hosting choice by itself supplies compliance.

Where does HyperLake fit, and when is it not the right choice?

HyperLake runs in the customer's own AWS or Google Cloud account today. Azure, OVH, private cloud and on-premises deployments are coming soon. Its FAQ describes managed data stacks, open models served through KServe and deployment-dependent integration scope.

Choose this approach when you need governed cloud-side evidence and repeatable data/model operations alongside a separately evaluated edge application. Choose a packaged industrial inspection system if its existing workflow meets the requirement with less integration. If all processing must stay on-site today, evaluate a currently available local stack rather than treating this architecture as an on-premises HyperLake offering.

How should a first pilot be bounded?

Choose one part family and one decision. Agree the defect taxonomy, independent test set, acceptable false accepts and rejects, operator workflow, retention and rollback before the trial. Record performance by defect class and operating condition, not just an average score.

Keep the initial test separate from the line's acceptance mechanism. Compare a locked model version against an independently reviewed image set, include normal production variation, and record every missed trigger or low-quality frame. Decide in advance who can change thresholds and who approves a new model. A limited observation run can expose capture and workflow problems before the system influences disposition.

Run a 2-week pilot. Agree scope, integrations and success criteria first; this wording names a proposed trial period, not a guarantee of production readiness in two weeks.

See the physical-AI hub, data infrastructure guide and robotics training-data guide. Compare the companion workflows: packaging surface-defect review and missing-component assembly review. The LLM cost calculator is useful only for an added language-model workflow; it does not size camera inference.

Frequently asked questions

Sources

1. Basler acA1920-40gc product specifications
2. NVIDIA DeepStream overview
3. NVIDIA Jetson Orin
4. EIA Table 2.4 — average electricity price by end-use sector, 2025
5. NIST Guide to Operational Technology Security
6. OSHA 1910.212 — General requirements for all machines
7. Office of the Privacy Commissioner of Canada — employee notice example
8. HyperLake FAQ
9. HyperLake use cases

Related guides

• Physical AI infrastructureWhat a Physical AI stack needs, why sovereignty matters for robot data and models, and how Hyperlake deploys and governs the stack in your environment.
• Data infrastructure for physical AI and robotics
• Robotics training-data formats and provenance