Senior Computer Vision Engineer - C++

About the CompanyKapdaa transforms UK garment waste into valuable recycled products. We are currently buildingAI4Fibres, an AI-powered textile recycling system which will run with UK textile waste.Role SummaryYou will build and own the real-time vision core of our sorting line: multiple synchronised camerastreams per lane, GPU inference, and a per-item decision that must be made before the garmentphysically reaches the diverter. This is a deep engineering role on the hot path C++, concurrency,memory and GPU throughput where the target is not accuracy alone but accuracy delivered insidea fixed time budget, continuously, for a full shift.You will work to throughput and latency budgets andcontribute to the architecture of the edge tier. This is a greenfield build on real hardware you willwork on the line itself, with the cameras, lighting and conveyor, alongside the ML team whosemodels you deploy. The throughput you achieve is not just your result: it sets the rate the rest of theplatform is designed around.Key Responsibilities●​ Image processing pipeline own everything upstream of inference: intrinsic/extrinsiccalibration and lens-distortion correction, homography from image space to the belt's encodercoordinate frame, flat-field and white/dark-reference correction, RGB↔hyperspectral spatialregistration, background subtraction and connected-component/morphological segmentationto extract per-item ROI.●​ Real-time C++ core multi-camera capture and processing under a strict per-item latencybudget: concurrency, thread and queue design, memory and buffer management, zero-copywhere it counts.●​ Multi-camera streaming synchronized capture across cameras and lanes, frame timing andalignment, handling dropped frames and degraded sensors without stalling the pipeline.●​ GPU inference integration CUDA/TensorRT, model export and optimisation (ONNX,quantisation), batching, and keeping the GPU fed rather than stalled.●​ Item identity & tracking assign identity at detection and carry it through classification torouting, using spatial/encoder-based tracking rather than fragile timing windows.●​ Performance engineering profile, measure and defend: know where every millisecondgoes, and produce numbers, not impressions.●​ Edge deployment & operations run reliably on GPU-backed edge devices (NVIDIA Jetsonclass) on the factory floor: startup, recovery, remote diagnostics, versioned rollouts, andbehaving sensibly when something upstream fails.●​ Vision pipeline quality work with the ML team on model integration, and build theinstrumentation that shows what the line is actually doing in production, not just what themodel scored in training.●​ Architecture contribution shape the edge tier's design with the team and hold the interfacecontract between the vision core and the rest of the platform.Required SkillsListed in order of expected depth expert command of the core, hands-on proficiency in the rest, andtooling you can pick up here.CORE EXPERTISE - expert depth requiredComputer vision & image processing fundamentals image formation, camera calibration and multi-camera registration, colour spaces and radiometric correction, segmentation and morphological operations, and the judgement to know when a classical technique beats a network on the hot path.Modern C++ (17+) and performance engineering multithreading and concurrency, lock and queue design, CPU and memory awareness, and real profiling experience (perf, Nsight, or equivalent). You have made a real system measurably faster and can explain exactly how.GPU inference in production CUDA and TensorRT, model conversion and optimisation, batching strategy, and debugging the pipeline when the model runs but the output is wrong.Multi-camera / multi-stream video pipelines GStreamer or DeepStream, OpenCV, industrial cameras (GigE Vision, GenICam a strong plus); synchronisation, buffering and backpressure across concurrent streams.Linux & edge strong fundamentals, Bash, and production operations on GPU-backed edge devices.WORKING PROFICIENCY - hands-on, used regularlyPython for tooling, evaluation and model work; PyTorch and ONNX exportgRPC + Protocol Buffers, message queues, streaming interfacesDocker, Git, GitHub Actions; scripted, repeatable deploymentReal-time system design fundamentals latency budgets, buffering, graceful degradationAI-assisted development: skilled with coding agents and LLM-based workflows (Claude Code,Codex) in everyday engineering without dropping the bar on correctness, tests orperformanceWAYS OF WORKING - how you operateMeasure, don't assume you quote latency and throughput numbers, and you can say howyou obtained them.High engineering standards code review, meaningful tests including load and latency testson the hot path, reliable CI/CD.Communication explains technical trade-offs clearly to engineers and to non-technicalcolleagues and in management.Start-up mindset fast-paced environment,