Collaboration, pair programming, and teamwork.

Taking ownership across the stack.

Test-driven development, and refactoring regularly to keep our codebases healthy.

Evolving the pipeline architecture so cameras can come and go, and configuration can change, without disrupting live video or losing object identity.

Setting the boundary between systems and models. You decide what belongs in the compiled service, what belongs in the inference graph, and where the performance ceiling really is.

Owning perception quality end to end. Defining what good looks like in numbers: evaluation data, a regression harness, and a defensible story on model choice for our hardware.

Extracting the maximum value from our sensors. Fusing every observation available while staying robust to occlusion, poor lighting, and deliberate evasion.

Taking the service from “runs” to “trustworthy unattended.” Failure detection, graceful degradation, and observability good enough that we know why something broke without a site visit.

Closing the loop from the field. Using deployment data to find headroom, and building the dataset and evaluation infrastructure that makes that repeatable.

Leading the work and the people. Setting direction across perception, reviewing the hard changes, and owning the interfaces perception exposes to the rest of the product.

Around 5+ years building production systems, with several at staff scope: owning a system’s architecture, not just its tickets.

Significant professional experience with perception or machine learning for hardware products in a safety-critical field - aerospace, robotics, medical devices, autonomous vehicles, or physical security.

Deep modern C++. This is a C++20 codebase with Abseil, gRPC, and CMake; you should be comfortable owning lifetime, threading, and shutdown semantics in a long-running daemon.

Real GStreamer or media-pipeline experience, DeepStream or another NVIDIA video stack: pads, probes, caps negotiation, bus messages, and the specific pathology of a pipeline that is alive but not moving.

Applied computer vision you have shipped - multi-object tracking, re-identification, or multi-camera association - and the judgment to know when the answer is a better model versus better geometry versus better plumbing.

A clear grasp of linear algebra, optimization, statistics, and algorithms, and the theory behind the techniques you reach for.

Experience across the deep-learning lifecycle: PyTorch or an equivalent framework, custom layers and operations, optimizing networks for inference on edge compute, reproducibility, and honest evaluation.

GPU inference in practice: TensorRT engines, batching, fp16, and reasoning about where latency actually goes.

Edge instincts. You have debugged something that only fails on the device, after nine hours, on one customer’s network, and you treat observability and failure classification as part of the feature.

Python fluency. A meaningful share of the load-bearing logic is Python, and you will be the one deciding what it costs us.

A generalist mindset - able to dive in wherever the bottleneck is, from cloud training infrastructure down to embedded systems.

Excellent written and verbal communication, and the ability to set technical direction and disagree productively with adjacent teams.

NVIDIA Jetson in production - JetPack, L4T, Yocto images, or the joy of cross-building for aarch64.

Previous experience building multi-camera tracking systems.

Video surveillance, VMS, or ONVIF/RTSP integrations, and knowing how cameras actually misbehave.

GPU architecture and CUDA programming.

Familiarity with VLMs and other multi-modal models for semantic scene understanding.

Owning model evaluation: datasets, metrics, and the discipline to reject a model that benchmarks better but ships worse.

Collaboration, pair programming, and teamwork.

Taking ownership across the stack.

Test-driven development, and refactoring regularly to keep our codebases healthy.

Evolving the pipeline architecture so cameras can come and go, and configuration can change, without disrupting live video or losing object identity.

Setting the boundary between systems and models. You decide what belongs in the compiled service, what belongs in the inference graph, and where the performance ceiling really is.

Owning perception quality end to end. Defining what good looks like in numbers: evaluation data, a regression harness, and a defensible story on model choice for our hardware.

Extracting the maximum value from our sensors. Fusing every observation available while staying robust to occlusion, poor lighting, and deliberate evasion.

Taking the service from “runs” to “trustworthy unattended.” Failure detection, graceful degradation, and observability good enough that we know why something broke without a site visit.

Closing the loop from the field. Using deployment data to find headroom, and building the dataset and evaluation infrastructure that makes that repeatable.

Leading the work and the people. Setting direction across perception, reviewing the hard changes, and owning the interfaces perception exposes to the rest of the product.

Around 5+ years building production systems, with several at staff scope: owning a system’s architecture, not just its tickets.

Significant professional experience with perception or machine learning for hardware products in a safety-critical field - aerospace, robotics, medical devices, autonomous vehicles, or physical security.

Deep modern C++. This is a C++20 codebase with Abseil, gRPC, and CMake; you should be comfortable owning lifetime, threading, and shutdown semantics in a long-running daemon.

Real GStreamer or media-pipeline experience, DeepStream or another NVIDIA video stack: pads, probes, caps negotiation, bus messages, and the specific pathology of a pipeline that is alive but not moving.

Applied computer vision you have shipped - multi-object tracking, re-identification, or multi-camera association - and the judgment to know when the answer is a better model versus better geometry versus better plumbing.

A clear grasp of linear algebra, optimization, statistics, and algorithms, and the theory behind the techniques you reach for.

Experience across the deep-learning lifecycle: PyTorch or an equivalent framework, custom layers and operations, optimizing networks for inference on edge compute, reproducibility, and honest evaluation.

GPU inference in practice: TensorRT engines, batching, fp16, and reasoning about where latency actually goes.

Edge instincts. You have debugged something that only fails on the device, after nine hours, on one customer’s network, and you treat observability and failure classification as part of the feature.

Python fluency. A meaningful share of the load-bearing logic is Python, and you will be the one deciding what it costs us.

A generalist mindset - able to dive in wherever the bottleneck is, from cloud training infrastructure down to embedded systems.

Excellent written and verbal communication, and the ability to set technical direction and disagree productively with adjacent teams.

NVIDIA Jetson in production - JetPack, L4T, Yocto images, or the joy of cross-building for aarch64.

Previous experience building multi-camera tracking systems.

Video surveillance, VMS, or ONVIF/RTSP integrations, and knowing how cameras actually misbehave.

GPU architecture and CUDA programming.

Familiarity with VLMs and other multi-modal models for semantic scene understanding.

Owning model evaluation: datasets, metrics, and the discipline to reject a model that benchmarks better but ships worse.