Define the Problem First
Before you post a job or shortlist candidates, write down what the computer vision system actually has to do — not the technology you think it should use. Answer these questions: what data will the model see (single camera, multi-camera, live streams, offline clips)? What must the output be (bounding boxes, keypoints, tracked identities, a score)? And what latency is genuinely required? A system that needs sub-30ms responses on edge hardware is a very different project from one that can afford a batch pipeline in the cloud.
Too many teams hire for a toolkit instead of a problem. If you want athlete pose analysis in live competition, you need someone who has shipped real-time keypoint detection, not just someone who has trained ImageNet classifiers. Nail the constraints first and the candidate shortlist becomes obvious.
Skills to Look For
Filter CVs against a short, concrete list rather than a wishlist of every model ever published. The skills that actually matter in production:
- Deep learning fluency — PyTorch and/or TensorFlow, plus confidence reading research papers and turning them into trainable code.
- Classical image processing — OpenCV fundamentals: camera calibration, geometric transforms, filtering, and efficient frame pipelines.
- Deployment and optimization — exporting models through ONNX Runtime or TensorRT, quantization, and hitting latency targets on real hardware.
- Tracking and multi-object systems — experience with YOLO, ByteTrack or DeepSORT when identities must survive occlusion.
- Production engineering — FastAPI services, Docker, cloud storage and compute (S3, EC2), and a habit of writing tests.
Questions to Ask
Skip trivia like "what is a convolution?" and ask questions that expose how someone thinks about trade-offs:
- Walk me through a model you took from prototype to production. What broke, and what did you change?
- Your pose model drops to 40ms on our edge device. How do you bring it under 30ms?
- How do you keep player identities stable when players occlude each other on every frame?
- What does your testing strategy look like for a vision pipeline that consumes live RTSP streams?
Good answers reference measurable work — latency numbers, accuracy improvements, coverage percentages — because those are the real currency of a production engineer.
Full-Time vs Contract
Contracting works when the scope is well-defined: a specific model, a latency target, a migration to TensorRT. You pay for outcomes and can walk away cleanly if the prototype misses. Full-time makes sense when the vision stack is your core product and you want one person owning it across many quarters — data pipelines, retraining, monitoring, and feature requests.
A pragmatic middle path is a paid pilot: two to four weeks, a real dataset from your domain, and a working demo with honest numbers. It tells you more than any interview loop about how the engineer communicates, estimates, and handles messy real-world data.
Evaluating a Portfolio
Ignore glossy screenshots. Ask to see the system running, then dig into the engineering. Check whether the portfolio reports metrics like inference latency, accuracy improvements, or test coverage — numbers imply the work survived real environments. Ask what the candidate would do differently next time; the answer reveals how much they actually owned the project.
For sports and motion work, the bar should be high: can they show pose estimation on live footage with occlusion, varied lighting and motion blur? Hassan's work at ID Sports Ventures covers exactly that — multi-object tracking through severe occlusion, a 15% pose accuracy boost across varied lighting, and sub-30ms edge inference on streaming footage.
What a Strong Candidate Looks Like
Beyond the checklist, the strongest candidates share a few habits. They ask about your data before your budget. They give straight answers about feasibility — including when something won't work. They can explain a trade-off between accuracy and latency in plain language. And they treat deployment as part of the job, not an afterthought.
"Hire for the ability to ship under real constraints. Demos are easy; production systems that hold their latency target on game day are not."