About the Role

We are seeking a Production Computer Vision & Inference Optimization Engineer to architect, scale, and refactor our core Vision Management System (VMS) and video analytics pipelines. Unlike research-only roles, this position focuses heavily on production inference efficiency, brownfield system refactoring, and GPU cost reduction.

You will take ownership of both building greenfield multi-stream video ingestion architectures and overhauling existing vision pipelines to maximize streams-per-GPU density without compromising detection accuracy.

What You’ll Do

- Design and build new multi-camera video ingestion and analytics systems from scratch.
- Refactor and modernize legacy VMS and vision pipeline architectures for performance and reliability.
- Optimize and quantize production models to reduce GPU usage and cloud/hardware running costs.
- Increase streams-per-GPU density while maintaining detection accuracy targets.
- Profile and resolve bottlenecks across decoding, memory transfer, inference, and post-processing.
- Deploy and scale concurrent models in production using inference serving infrastructure.
- Partner with product, ML, and DevOps teams to define benchmarks and cost-per-stream targets.

Key Requirements

- Model Compression & Quantization: Proven production experience converting vision models (YOLO series, RT-DETR, Transformers, segmentation) from FP32 to FP16 and INT8 using NVIDIA TensorRT and ONNX Runtime, utilizing Post-Training Quantization (PTQ) calibration and Quantization-Aware Training (QAT) to prevent accuracy drop.
- High-Throughput Video Pipelines: Hands-on expertise building and debugging low-latency, multi-camera RTSP/WebRTC video ingestion pipelines using NVIDIA DeepStream, GStreamer, or FFmpeg,



leveraging hardware-accelerated decoding (NVDEC) and zero-copy GPU memory management.
- Inference Serving & GPU Profiling: Experience deploying concurrent models via NVIDIA Triton Inference Server (dynamic batching, concurrent execution) and profiling GPU bottlenecks (VRAM, compute utilization, PCIe bandwidth) in Python and C++/CUDA environments.
- Production Mindset: A track record of shipping and maintaining real-time vision systems in production, not only training models in notebooks.

Nice to Have

- Experience with edge deployments (e.g., NVIDIA Jetson Orin) and GPUs such as T4, L4, or A10.
- OpenCV (CUDA), model pruning, and GPU cost/FinOps experience.
- Familiarity with VMS / CCTV platforms and camera integration standards.
- Experience with Kubernetes, Docker, and cloud GPU infrastructure.

Tools & Technologies TensorRT · ONNX Runtime · NVIDIA DeepStream · GStreamer · FFmpeg · RTSP · WebRTC · NVDEC · Triton Inference Server · CUDA · C++ · Python · OpenCV

About Us

Super AI Polaris builds sovereign AI platforms designed for education systems, enterprises, and governments. Our solutions combine advanced artificial intelligence with practical deployment models to solve real-world challenges at scale.

Equal Opportunity

Super AI is an equal prospect employer. We welcome applicants of all backgrounds and do not discriminate on the basis of gender, religion, caste, age, disability, or any other protected characteristic.

How to Apply

Send your resume to [email protected] with the subject line “Application – Senior Computer Vision & GPU Optimization Engineer”. Please include links to your GitHub, portfolio, or relevant project work where applicable.

Pay: ₹360,000.00 - ₹960,000.00 per year

Benefits

- Paid sick time
- Paid time off

Work Location: In person

📌 Senior Computer Vision & GPU Optimization Engineer (Delhi)
🏢 Super AI Polaris
📍 Delhi