1. Own Video Intelligence
- Build and train CV models for driver fatigue & distraction detection, ADAS-style road & event detection, and cargo, theft, and in-cabin monitoring.
- Turn messy, real-world video into reliable detections.
2. Optimize for the Edge
- Make models run cost-effectively at scale using quantization, pruning, distillation, on-device/edge inference, and trigger-based, event-driven processing.
- Treat inference cost-per-camera as a first-class design constraint.
3. Train, Don't Just Wrap
- Build custom models where they create differentiation.
- Use pre-trained backbones and transfer learning to move fast.
- Know when to fine-tune vs. build from scratch.
4. Own the Vision Data Pipeline
- Define annotation specs and quality standards (labeling is outsourced — you own the spec).
- Build training and evaluation datasets from real fleet video.
- Monitor model drift and retrain as conditions change.
5. Ship to Production
- Deploy models into the product, not notebooks.
- Build inference services (edge + cloud), monitoring, and versioning.
- Iterate from real field performance.
6. Collaborate Across Teams
- Work with Hardware/IoT Engineers on dashcams and edge devices.
- Partner with Data & AI Product Engineers for shared data and benchmarking.
- Collaborate with Software Engineers and Product/Leadership to integrate solutions and refine use cases.
Must-Have
- Strong computer-vision and deep-learning fundamentals (object detection, image/video models)
- Hands-on with PyTorch or TensorFlow — training, not just inference
- Track record deploying CV models to production (real users, real data — not just papers or Kaggle)
- Experience optimizing models for real-time / resource-constrained inference
- Solid engineering (Python; can build and ship services)
- Comfort with messy, real-world image/video data at scale
Nice-to-Have
- Edge / embedded deployment (NVIDIA Jetson, mobile, on-device, TensorRT/ONNX)
- Driver monitoring / ADAS / dashcam / automotive vision experience
- Data-centric ML and annotation-pipeline design
- Inference cost optimization at fleet scale
- MLOps: model versioning, monitoring, automated retraining
