What you will own

Detection and tracking models that ground higher-level reasoning
VLM selection, adaptation, and fine-tuning for security incident understanding
Temporal / multi-frame modeling so the system reasons over clips, not just frames
Image-quality handling: ROI-aware processing, low-light enhancement, denoise, super-resolution, and keyframe selection
Distillation of a lightweight first-pass model so the system stays accurate and cheap at scale
Tight loop with evals: hard negatives, false positives / negatives, and operator overrides into model improvements

What we are looking for

Strong experience in computer vision, multimodal ML, or video understanding
Shipped models that had to work on messy real-world data, not just curated benchmarks
Comfortable moving across detection, tracking, representation learning, VLMs, and model optimization
Pragmatic about balancing quality, latency, and cost in production
Production-quality Python and comfort with training pipelines, experimentation code, and model evaluation tooling