Auriga - Lead Engineer - Artificial Intelligence

Description : Position Overview : Were hiring an AI Lead Engineer to architect, ship, and scale production-grade Computer Vision with a focused GenAI charter.Youll lead 610 ML/CV engineers to deliver high-accuracy, low-latency video analyticsdetection, tracking, segmentation, recognition, re-identificationtackling multi-camera tracking, 24/7 streaming at scale, long-tail drift, and edge optimization on GPUs/Jetson, while partnering with Product/Platform to meet FPS, latency, accuracy, and cost SLAs.The day-to-day centers on core CV model design, training/leval, deployment, and streaming/edge performance, complemented by GenAI that amplifies the stack : integrating VLM/LLM capabilities for natural-language video search, incident summarization, operator Q&A, and copilot workflows; standing up RAG over video/sensor/metadata embeddings with robust prompts, tooling, evals, and guardrails; and driving data ops with synthetic data, auto-labeling, active-learning triage with privacy, safety, and cost controls.What Youll Own : End-to-end delivery of CV products : problem framing / data/labeling / model design / optimization / deployment/ monitoring / iteration.Technical roadmap & architecture for video analytics pipelines (ingest / decode / infer / track / post-process / store/serve).Team leadership : mentoring, hiring input, OKRs, code/research standards, and performance coaching.Key Responsibilities : Leadership & Execution : - Translate business goals (e.g., reduce shrinkage, increase throughput, improve safety) into measurable CV objectives and SLAs/SLOs (e.g., mAP/IDF1, per-frame latency, dropped-frame rate, cost/stream).- Lead design reviews; establish MLOps and coding standards; enforce experiment tracking, reproducibility, and dataset/version governance.- Drive capacity planning, GPU/Jetson utilization, batching/windowing strategy, autoscaling, and cost Vision R&D (Detection, Tracking, Segmentation, Recognition) : - Deliver production models for object (e.g., YOLOv8/v9,Mask R-CNN/Mask2Former, EfficientNet/ConvNeXt, Build person/vehicle/product re-ID, face/attribute recognition (e.g., ArcFace/CosFace), OCR (e.g., PP-OCR), and keypoint/action recognition (e.g., MMPose, SlowFast/X3D).- Tackle domain adaptation, class imbalance, and occlusions; design augmentations and semi-supervised/active learning loops to harvest hard Analytics & Edge Inference : - Architect real-time pipelines using NVIDIA DeepStream/GStreamer/OpenCV; optimize decode (NVDEC), pre/post, and trackers for 3060 FPS at 1080p.- Optimize with TensorRT/ONNX Runtime/Torch-TensorRT, INT8 calibration, pruning/distillation; leverage Jetson Orin/Xavier/Nano and DLA where applicable.- Design multi-camera fusion, homography/camera calibration, and cross-camera ID consistency for retail, traffic, manufacturing, and security use cases.- Implement privacy-by-design features (e.g., face/license blur, PII AI & LLMs : - Architect robust RAG : retrieval pipelines with Pinecone/ChromaDB/Milvus; index sharding/compaction; freshness policies; hybrid search.- Design agents with LangChain/LangGraph; implement tool-use, safety filters, and guardrails; add evaluation loops (e.g., Serving & MLOps : - Ship services via FastAPI/Flask; containerize with Docker; orchestrate on Kubernetes (KServe) or AWS SageMaker/Vertex AI.- Build high-throughput inference with Triton Inference Server (dynamic batching, concurrent models, model ensembles).- Streaming & storage : RTSP/RTMP ingest, Kafka/Kinesis, object storage + time-series DB; index/frame-level metadata for search and analytics.- CI/CD with MLflow/DVC (artifacts, model registry), unit/integration tests, and rollout strategies (canary, Drift & Governance : - Production monitoring with Prometheus/Grafana; per-stage latency, FPS, GPU memory/SM occupancy, dropped frames, and backpressure.- Model observability : data/feature drift, concept drift on detections/tracks, re-ID distribution shifts, outlier/novelty detection, safety metrics.- Human-in-the-loop review tools (CVAT/Label Studio) and auto-retraining triggers; maintain model cards, evaluation reports, versioned prompts/configs, and auditability.- Ensure compliance and privacy/PII handling; ONVIF/edge security best & People Leadership : - Partner with Product/SRE/DevOps on roadmaps, SLAs, incident response runbooks, and cost/perf tradeoffs.- Lead and grow a 610 person CV team; foster a culture of high-quality experiments, rigorous reviews, and measurable impact.- Communicate progress/risks to executives with clear, metric-driven updates and customer-facing Qualifications : - Candidate shall have a degree in B.E/B.Tech/MCA in any discipline preferable computer science- 5-8 years in ML/Computer Vision with 2+ years leading 610 engineers delivering production video analytics.- Proven track record shipping systems with business impact (accuracy, latency, cost).Strong in at least one per category (and comfortable across most) : - CV Frameworks : PyTorch (preferred) or TensorFlow; OpenCV, NVIDIA : Triton Inference Server, FastAPI/Flask.- Optimization : TensorRT, ONNX Runtime, Torch-TensorRT; quantization/pruning/distillation; INT8 calibration.- Tracking/Re-ID/OCR : ByteTrack/OC-SORT/DeepSORT; ArcFace/CosFace; PP-OCR/Tesseract.- Agents & Retrieval : LangChain or LangGraph; MLOps : Docker, Kubernetes (KServe/SageMaker/Vertex AI), MLflow/DVC.- Cloud : AWS (SageMaker, EC2/EKS) or GCP (Vertex AI) or Azure ML (AKS).- Programming : Python (expert); C++ or Go for perf-critical components; CUDA fundamentals a plus.- Streaming/IO : RTSP/RTMP, Kafka/Kinesis/Rabbitmq; ONVIF familiarity.- Strong system design (multi-stream pipelines, GPU scheduling, distributed tracking/indexing) and excellent Qualifications : - Operating vector or time-series stores for video metadata at 10M50M+ rows; search over tracks, IDs, and events.- Experience with multi-camera tracking, calibration and zone-based analytics.Jetson fleet management : - Ray/Ray Serve or Kubeflow; feature stores (Feast); complex event processing.- Domain experience in one or more : retail analytics, traffic/ADAS, manufacturing QA, security/safety, sports analytics.- Open-source contributions, patents, or publications in CV/video analytics.(Nice to have) Multimodal exposureCLIP/SigLIP, SAM/Mask2Former, or VLMs for captioning/searchused sparingly to support CV workflows (ref:hirist.tech)