województwo mazowieckie, Polska
Onsite
The video understanding pipeline—from multi-object tracking and re-identification to action recognition—is yours to own and scale. The technical challenge lies in balancing high-throughput, low-latency GPU inference with cutting-edge detection accuracy across massive streams of sports footage. You will establish rigorous benchmark suites and decide which vision foundation models make it into production.
The platform ingests live and recorded sports video to automatically extract real-time metadata, temporal key moments, player statistics, and scene-boundary events. Operating across multiple sports leagues, it relies on frame-accurate video parsing to power downstream broadcast automation, search, and analytics tooling.
Technology Stack: The core ML pipeline is built in Python using PyTorch for model training and fine-tuning. Production GPU inference is optimized for cost and latency using TensorRT and ONNX Runtime within Dockerized, queue-driven microservices. FFmpeg handles large-scale video decoding, while Azure (Service Bus, Blob Storage) underpins the asynchronous messaging infrastructure.
Own the video understanding pipeline, including object detection, multi-object tracking, player re-identification, jersey-number recognition and on-screen text recognition (OCR).
Own event and action recognition, including temporal localisation of key moments, shot and scene-boundary detection, replay detection and camera-view classification.
Optimise GPU inference for throughput and cost through batching, precision tuning and compilation.
Monitor production quality and latency, and investigate failures and regressions.
Build and maintain labelled sports benchmarks, split by sport, league and content type.
Define and track key metrics, including mAP, F1, precision, recall, IoU, latency and GPU cost.
Run systematic bake-offs comparing new detection, tracking, segmentation, re-identification and video foundation models. Reproduce published results on our data.
Write decision memos with clear recommendations, trade-offs and licensing considerations.
Own the regression suite and shadow or A/B testing pipelines, ensuring every model or configuration change is validated before deployment.
Track new papers, model releases and leaderboards, and identify which developments warrant further evaluation.
3–5+ years of experience shipping production ML systems, including 2+ years in computer vision, ideally working with video.
Strong Python and PyTorch skills, including hands-on experience profiling and optimising GPU inference using TensorRT and ONNX Runtime.
Hands-on experience in at least two of the following areas: object detection and segmentation (YOLO, DETR-family, SAM), OCR and scene-text recognition, or video action recognition and temporal localisation.
Experience building test sets, calculating evaluation metrics such as mAP, F1 and temporal IoU, and running rigorous model comparisons.
Ability to critically evaluate new research papers and model cards, and reproduce published results.
Production deployment experience with Docker, queue-driven services and cloud GPU infrastructure.
Experience with video processing at scale, including FFmpeg, video decoding pipelines and frame sampling.
Strong written communication skills, with the ability to document technical trade-offs and make clear recommendations.
Experience with Azure (Service Bus, Blob Storage) and message-driven architectures.
Experience with vision-language models and multimodal understanding across video, audio and text.
Experience fine-tuning or adapting vision models to specific domains, particularly sports video.
Experience with self-supervised learning or synthetic data generation to reduce labelling costs.
Familiarity with annotation tools such as CVAT and Label Studio, and experience working with labelling teams.
Experience with shadow deployments and A/B testing.
Familiarity with commercial model licensing.
Direct, unconstrained ownership of the full ML lifecycle—from paper replication and benchmark creation to live GPU optimization.
Production impact at scale, where your model optimization directly lowers real-world infrastructure costs and improves latency on high-volume video data.
High-rigor, evidence-driven engineering environment that prioritizes A/B testing, shadow paths, and statistical validation over hype.
