Dedicated Computer Vision / ML Delivery Team (Monthly Retainer)

Description We run a computer vision and ML agency on Upwork with a steady pipeline of client projects across sports, retail, medical, industrial, and general video and image analytics. We want a small Pakistan based development team to become our dedicated delivery team on a flat monthly retainer. This is a long term relationship, not a one off gig. You become our development back office. You handle the technical build across multiple concurrent projects while we handle sales, client relationships, and delivery direction. You work as a subcontractor to our agency and present to clients as part of our development team. What you deliver, across several projects at once: Object detection with RF-DETR, RTMDet, and YOLO. Multi object tracking. Pose estimation on video. Training and fine tuning on RunPod cloud GPUs. Dockerized, reproducible training and inference pipelines. Dataset preparation and cleaning. Evaluation against defined accuracy targets. Occasional Power BI dashboards and Next.js glue to wire a model into a client facing app. You need strong spoken and written English. You join client video calls as part of our development team, so you must hold a technical conversation on camera with confidence. How the engagement works: Flat monthly retainer, starting around USD 700 per month, covering an agreed capacity of concurrent work. As project volume grows we raise the retainer and expand the team with you. We value reliability, real capacity, and overlap hours over price. You sign an NDA. All IP assigns to our agency. Several of our engagements are IP to client with client NDAs in place, so the assignment and confidentiality terms are non negotiable and go in place before any work touches client data. We want an exceptional small team we can grow with, not the cheapest bidder and not a reseller. If you have real shipped CV and ML work, readable code, GPU training experience, and English that can face a client, apply. To apply, answer the screening questions below. Applications that skip them or send only demo videos will not be reviewed. Responsibilities Build and ship object detection across projects using RF-DETR, RTMDet, and YOLO Implement multi object tracking and pose estimation on video Train and fine tune models on RunPod cloud GPUs Package training and inference into Dockerized, reproducible pipelines Prepare, clean, and version datasets for training and evaluation Evaluate models against defined accuracy targets and report results clearly Build occasional Power BI dashboards and Next.js glue to deliver models into client apps Run several concurrent projects in parallel without dropping delivery Join client video calls and represent as part of our development team Work under NDA and assign all IP to our agency Required skills Python and PyTorch to production standard Object detection with RF-DETR, RTMDet, and YOLO Multi object tracking Pose estimation on video RunPod or equivalent cloud GPU training and fine tuning Docker and reproducible training and inference pipelines Dataset preparation, cleaning, and evaluation to accuracy targets Some Power BI and Next.js for delivery glue Strong spoken and written English able to face clients on video calls Small team with real concurrent project capacity Nice to have Domain experience in sports, retail, medical, or industrial video analytics Experience owning a full pipeline from raw footage to a deployed model Prior work as a subcontractor or white label development team MLflow, Weights and Biases, or similar experiment tracking Roboflow or similar annotation and dataset tooling Familiarity with model export and optimization such as ONNX or TensorRT Experience level Expert. Established team with a real shipped CV and ML portfolio and hands on GPU training experience Screening questions Link two or three CV or ML projects you shipped, each with a public repo or code sample we can read. For one of them, name the model architecture, the framework, and the file where the core training or inference loop lives. Do not send demo videos in place of code. How many people are on your team, what does each person do, and how many concurrent projects can you actively deliver right now while keeping each on schedule? Walk through your last real RunPod training run. Which GPU and pod type, what was in the Docker image, how did you get data onto the pod, and what exactly makes the run reproducible if we hand it to another engineer. Describe a detection or pose model you trained to a defined accuracy target. State the metric, the target number, the number you hit, the dataset size, and how you ran evaluation. Record a one to two minute intro video of the exact team member who would join our client calls, speaking to camera about one of the projects above. We are assessing spoken English, so a scripted voiceover will not pass. Are you willing to sign an NDA and assign all IP to our agency before any client work begins? Answer yes or no. State your working hours in UTC, your overlap window with the Americas and Europe, and your typical response time on an active project.