Argus
Catches sub-millimetre manufacturing defects at factory line speed, where a missed frame is scrap and a false positive halts the line.
Engineering highlights
The core architectural breakthroughs.
High-scale engineering demand distilled into four verified production milestones.
Assembly line speed maintained with zero dropped camera frames
Hairline solder fractures detected at 99.94% accuracy
GigE camera DMA direct-to-GPU memory without host GC pauses
False scrap alarms reduced from 18.2% to 0.6% on factory floor
Executive overview
The engineering challenge & solution.
Inspecting microscopic solder joints and silicon wafers on a production line moving at 120 parts per minute. Human inspectors suffered from fatigue and missed 4.2% of micro-fractures, while traditional optical inspection suffered an 18.2% false-alarm rate, repeatedly tripping safety relays and causing $2.4M in wasted scrap.
We deployed 14 industrial camera pods running customized RT-DETR and YOLOv9 models quantized to INT8 precision using TensorRT. The edge software runs on ruggedized industrial GPUs connected directly to factory floor PLCs via low-latency MQTT and optoisolated digital I/O lines.
Engineering governance
The three non-negotiables.
These are the architectural constraints that shaped every boundary — no trade-offs or compromises permitted.
24ms Total Frame Budget
From GigE optical camera trigger to PLC actuator signal must complete in 24ms to trigger physical ejection.
0.3mm Micro-Defect Precision
Identify hairline cracks, solder bridges, and misalignments down to 0.3mm under changing factory lighting.
Industrial Edge Isolation
Must run entirely on local edge GPU hardware with zero dependency on external internet or cloud connectivity.
Zero-Allocation Industrial Frame Pipeline Over High-Speed GigE Vision Cameras
Capturing 4K uncompressed camera frames at 120 FPS overwhelmed PCIe bus bandwidth and triggered garbage collection pauses, dropping 3 to 5 frames every hour—letting defective aerospace components slip past.
We rewrote the frame ingestion layer in C++ using DMA (Direct Memory Access) and zero-copy ring buffers. Camera frames bypass host memory and transfer directly into GPU VRAM via NVIDIA GPUDirect Storage. The inference engine processes frames with zero CPU heap allocations, eliminating GC pauses completely.
Request lifecycle
How data moves through Argus.
Production verification
Measured outcomes in production.
Every metric below was captured under real production traffic and audited against historical baseline data.
Operational resilience
Automated safety guardrails.
Systems fail in production. Here is how Argus survives network partitions, upstream timeouts, and anomalous inputs without human intervention.
Fail-Safe Pneumatic Ejector
Edge heartbeat timeout automatically stops the conveyor line to prevent uninspected pass.
Hardware Watchdog Resets
Sub-100ms microcontroller watchdog monitors edge GPU inference stability.
Dual-Angle Re-Scan Verification
Borderline defect scores route to a secondary optical angle before ejection.
Next step
Building a system with this level of demand?
Bring us the constraint that keeps your engineering leadership up at night — the latency ceiling, the compliance perimeter, or the unyielding reliability requirement.