Bleeding Edge Tech

Hello,

I’ve built a face-based event registration and attendance system using InsightFace (GPU), where users self-register via camera. I’ve also implemented an agent to automate onboarding and attendance workflows.

I now want to move beyond a POC into a technically challenging problem with real-world complexity.

Some directions I’ve explored:

• Moving from image-based (2D) recognition to video-based (temporal / 3D understanding)

• Robustness under real-world conditions (lighting, occlusion, motion blur)

• Real-time multi-camera identity tracking across a venue

• Using LLMs to analyze event data (engagement, behavior patterns, etc.)

However, these feel like incremental extensions.

What I’m really looking for is a hard, open problem at the intersection of vision systems, real-time inference, and agentic AI —something where current approaches break down in practice (not just benchmarks).

For example:

• Where do current face recognition / tracking systems fail at scale in real deployments?

• Are there unsolved challenges in combining vision models with LLM-based agents for real-world decision-making?

• Any known gaps between research and production systems in this space?

Would appreciate pointers to concrete, technically challenging problems worth tackling.

Thanks