SentientSight | Edge AI & Embedded Vision

Embedded vision in 2026 is not the Jetson TK1 era. The platforms, the runtimes and the models have all turned over, and hardware lifecycles are now a first-order engineering concern. SentientSight designs, builds and ships edge computer-vision systems, with OpenCV doing the image work and the right accelerator runtime doing the inference. See also our edge platform post.

Platform selection

NVIDIA Jetson Orin (Nano / NX / AGX). The default choice for GPU-accelerated edge vision: CUDA, TensorRT and a mature GStreamer camera stack on JetPack. Orin Nano is the budget entry point; Orin NX and AGX Orin scale up for multi-camera and multi-model workloads.
NVIDIA Jetson Thor. The Blackwell-based successor, now shipping, aimed at humanoid and general robotics and at workloads that mix vision with on-device generative AI. JetPack 7 supports both Orin and Thor.
Raspberry Pi 5 + AI Kit / AI HAT+. A Hailo NPU on a board most teams already know. Excellent price/performance for single-camera detection and classification when you do not need CUDA.
NPU-accelerated SBCs. Rockchip, Qualcomm and Intel-based boards with integrated NPUs, targeted through OpenVINO, LiteRT or vendor SDKs; OpenCV 5's new hardware acceleration layer (Arm KleidiCV, Qualcomm FastCV) speeds up the pre- and post-processing on the CPU side.

We benchmark candidate boards against your actual model and camera configuration before you commit to a bill of materials.

Planning around the 2026 Jetson EOL wave

In April 2026 NVIDIA pulled forward end-of-life for several LPDDR4-based Jetson modules, driven by memory pricing and supply. If your product has a 5-10 year field life, the module it ships on has to be chosen with the lifecycle roadmap in view, not just the benchmark. We help you:

Map current and announced last-time-buy dates against your production schedule.
Abstract the application from the module so a carrier-board or module swap is a rebuild, not a rewrite (containerized JetPack deployments, pinned TensorRT engine builds per target, hardware-abstracted camera capture).
Qualify a second-source platform early, while the first one is still shipping.

Runtime choices

TensorRT on Jetson for the lowest latency on NVIDIA GPUs; we build engines per module and per JetPack version, and cache them.
ONNX Runtime as the portable default across x86, Arm and NPUs via execution providers.
OpenVINO for Intel CPUs, iGPUs and NPUs.
LiteRT (formerly TensorFlow Lite) for Android, microcontroller-class targets and NPU delegates.
OpenCV 5's DNN engine when you want one dependency: 80%+ ONNX operator coverage, CPU-only for now, with the classic engine still available for CUDA and OpenVINO backends.

Model choices

Real-time detection in 2026 usually means the YOLO family — currently YOLO26, which is NMS-free end to end, which simplifies the post-processing you have to write on a constrained device. We distill, prune and quantize (FP16, INT8 with calibration, QDQ graphs) to hit your latency budget, and validate accuracy on your data after each step, because quantization that looks fine on COCO can fall apart on a factory line.

Camera and ISP integration

OpenCV is still the glue: GStreamer and V4L2 capture, Bayer and ISP handling, undistortion and rectification from OpenCV 5's calib and stereo modules, color-space conversion, ROI cropping and resizing in front of the model, and drawing and encoding behind it. Zero-copy handoff between the camera, the CPU and the GPU is where most edge pipelines lose their frame budget; we design for it from the start.

Typical engagements

Inspection line. Multi-camera defect detection at line speed on Orin NX, with a TensorRT YOLO26 model and OpenCV handling triggering, registration and measurement.
Retail analytics. Privacy-preserving people counting and dwell analysis on Raspberry Pi 5 + AI HAT+, aggregated on device, with no video leaving the store.
Robotics. Stereo depth, visual odometry and object detection fused on Jetson Thor, with lifecycle-aware module selection for a long-lived platform.

Edge feasibility study

Not sure which board, runtime or model will meet your budget and latency targets? Our fixed-scope feasibility study benchmarks your model on two or three candidate platforms with your camera setup and delivers a written recommendation, including the lifecycle risk for each option.