Computer Vision Engineer

Mayuresh More

Mumbai, India · +91-996*-***-*** · **************@*****.*** · LinkedIn · GitHub

Computer Vision Engineer with 2+ years building and deploying production CV systems. Core expertise in real-time video pipelines (RTSP, GStreamer, FFmpeg, OpenCV), object detection and tracking

(YOLOv8/YOLO11), and edge inference optimization with TensorRT on NVIDIA Jetson. Experienced training custom detectors, converting and deploying models across multiple hardware targets (Jetson Orin, RK3588, Raspberry Pi), and delivering end-to-end production CV systems on constrained hardware. Technical Skills

Vision & CV YOLOv8 / YOLO11, OpenCV, GStreamer, FFmpeg, NVIDIA DeepStream, object detection

& tracking

Video Pipelines RTSP multi-camera ingestion, real-time inference, WebSocket/HTTP streaming, multi-threading

Edge & Inference TensorRT (Jetson Orin), ONNX, TFLite, CoreML, RKNN (RK3588), model quantization & pruning

ML / Training PyTorch, YOLOv8/YOLO11 fine-tuning, custom dataset curation, annotation workflows Backend & Infra Python, FastAPI, Docker, Linux (Ubuntu), AWS (S3, Lambda, SQS, Batch), PostgreSQL Observability Structured logging, watchdog timers, performance profiling, latency benchmarking Work Experience

Computer Vision & Backend Engineer · Wiserli · Remote Nov 2025 – Present Production VSaaS platform · real-time AI video analytics · Python · TensorRT · NVIDIA Jetson · FastAPI · Docker · Linux

– Optimized YOLO model inference on NVIDIA Jetson Orin using TensorRT — profiled GPU/CPU utilization, resolved bottlenecks, and achieved real-time performance on a production always-on video analytics platform via remote SSH-based debugging and iterative tuning.

– Extended and maintained a production multi-threaded real-time video analytics pipeline processing live RTSP camera streams — refactored frame capture (FreshestFrame), GPU inference workers, and HTTP/WebSocket streaming layers to ship new analytics features without disrupting live client traffic.

– Hardened RTSP stream reliability on a production always-on platform by redesigning automatic reconnection logic, thread-safe frame buffers, and watchdog timers — eliminating the stream dropout issues that were causing client-facing downtime.

– Architected scalable asynchronous pipelines using Celery, RabbitMQ, and Redis to automate multi-format model conversion (PyTorch ONNX / TensorRT / TFLite) with encrypted verification and Supabase storage.

– Containerized a multi-service Jetson application (TensorRT inference backend + React/SocketIO frontend) using Docker Compose for seamless migration between Jetson devices — resolved JetPack/CUDA dependency constraints and engine portability issues to enable zero-downtime device swaps in production. Computer Vision & ML Engineer · Witsense AI · Remote Mar 2024 – Sep 2025 Custom object detector training · auto-annotation platform · AWS MLOps · edge optimization

– Trained custom YOLOv8/YOLO11 detectors achieving 98.5 mAP@50 across domain-specific targets

(door handles, top-view people, uncommon objects) — handled full pipeline from raw data collection and annotation through fine-tuning and evaluation.

– Optimized trained models for constrained edge hardware — achieved 80% model size reduction and 5 faster inference on Raspberry Pi 5 + Coral USB through INT8 quantization and architecture pruning.

– Engineered an end-to-end auto-annotation pipeline (GroundingDINO + Grounded-SAM) that converted raw images into production-ready YOLO datasets automatically — significantly cutting manual labelling effort on the platform.

– Architected serverless AWS MLOps infrastructure (SQS Lambda Batch S3) for GPU training job orchestration, reducing deployment latency by 40% on the annotation platform.

– Built and maintained FastAPI + PostgreSQL APIs for image upload management, annotation job tracking, and automated dataset validation pipelines.

Freelance Computer Vision Engineer · Independent Consulting · Remote 2025 Real-time multi-camera CV surveillance · jewellery manufacturing facility · edge deployment on RK3588

– Designed and deployed a production 4-camera real-time surveillance system processing 2K-resolution RTSP streams on Vicharak Axon RK3588 — leveraged GStreamer hardware-accelerated decoding to achieve low-latency on-device analytics without cloud dependency.

– Converted YOLOv8n to native RKNN format for RK3588 NPU inference — enabling real-time detection, tracking, and ROI-based guard presence monitoring within high-value asset zones (gold handling areas).

– Built FastAPI + PostgreSQL multi-camera ingestion APIs with WebSocket-based live alerting for policy violations; delivered complete POC with on-site data storage, reconnection logic (GStreamer + OpenCV), and full deployment documentation.

Projects

– Auto-rickshaw Detector — End-to-end YOLOv8 object detection model: self-collected and manually annotated a custom dataset, trained and evaluated the model from scratch. github.com/MayureshM0re/Yolov-8-auto-rickshaw-detection

– Face Recognition Access Control System — Built an employee/guest recognition pipeline using SCRFD for fast face detection and ArcFace for identity embedding and matching. Implemented known-person registration, real-time inference on live camera feed, and threshold-based guest vs employee classification. Education

B.E. Computer Science · University of Mumbai GPA 8.04 / 10 · Jul 2023 Hackathons & Certifications

Participant, Hugging Face LeRobot Worldwide Hackathon (2025) · NVIDIA: Simulating Your First Robot in Isaac Sim.

Contact this candidate