Работа в Москве, вакансия - Computer Vision & Real-Time Rendering Engineer • Москва - id409

Nextologies is an international company specializing in telecommunications and video content delivery and processing solutions. As we expand our team, we are looking for a Computer Vision & 3D Engineer for full time or contract based position.

About the role

We' re building a real-time visual replacement pipeline that operates at broadcast frame rates (60fps+). The system ingests live video, performs geometric scene understanding, reconstructs portions of the scene using modern neural rendering techniques, and composites replacement content seamlessly — all in real time.

This is a deep technical role at the intersection of computer vision, neural scene representation, and GPU-accelerated rendering. You' ll be working on production-grade infrastructure where latency and visual quality are both non-negotiable.

What you' ll work on

• Camera pose estimation and geometric localization from video frames

• Neural scene reconstruction using Gaussian splatting (3DGS / 4DGS) — including capture, training, and real-time inference pipelines

• Real-time inpainting and compositing on reconstructed scene geometry

• GPU pipeline optimisation for sub-frame latency on NVIDIA hardware

• Training data generation for supporting ML models (synthetic image generation, perspective transforms, augmentation pipelines)

• Integration with broadcast video infrastructure (SDI, NDI, or similar)

• Building and maintaining production services on Linux — deployment, monitoring, profiling, and debugging in headless GPU environments

Required experience

• Strong background in computer vision — camera models, homography, PnP, pose estimation

• Hands-on experience with Gaussian splatting pipelines (3DGS, 4DGS, or equivalent) — not just academic familiarity

• Proficient in Python and C++/CUDA — writing and optimising custom CUDA kernels, not just calling library APIs

• Deep familiarity with the NVIDIA stack: CUDA, cuDNN, TensorRT, NCCL, and associated tooling (Nsight, nvprof)

• Experience with Linux-native GPU workflows — driver management, CUDA environment configuration, headless rendering (EGL/OSMesa), containerisation with GPU passthrough (Docker + nvidia-container-toolkit)

• Experience optimising inference pipelines for hard real-time constraints

• Proficient with PyTorch including custom autograd, torch.compile, and mixed-precision workflows

• Comfortable working entirely in Linux environments — no Windows/Mac fallback

Highly desirable

• Experience with Unity 3D or Unreal Engine for real-time rendering and compositing — particularly headless or server-side rendering on Linux

• Background in broadcast or live production environments

• Experience with synthetic training data generation — perspective simulation, image augmentation, annotation pipelines

• Familiarity with object detection / segmentation models and their limitations at broadcast frame rates

• Experience with video streaming and low-latency pipelines — GStreamer, FFmpeg, or similar, ideally with hardware-accelerated encode/decode (NVENC/NVDEC)

• Profiling and tuning GPU memory bandwidth — understanding occupancy, warp divergence, and memory coalescing at a practical level

• Experience with multi-GPU setups and pipeline parallelism

• Background in sports broadcast or live event production technology

You might be a fit if you' ve...

• Built or contributed to a Gaussian splatting capture-to-inference pipeline end-to-end

• Shipped something that runs on GPU in a hard real-time constraint

• Done any kind of scene reconstruction or AR overlay work in a live video context

• Worked on training data pipelines for vision models at scale

• Debugged a CUDA pipeline in a headless Linux environment under production pressure

• If you think you can do it better, we are listening…

What we' re not looking for

Standard YOLO/feature-matching solutions and pixel-level replacement approaches are well-understood territory — we' ve specifically moved past them. We need someone who understands why geometry-based and neural rendering approaches outperform them in this context, not someone who needs convincing.