When every frame fights back, this is how AI survives

What happens when AI meets chaotic traffic, low-grade cameras, and unpredictable weather? It stops being theory and starts fighting to survive. Inside the real world of video AI where every frame is a battlefield and only the robust make it out

- Harsh Sharma

Artificial Intelligence (AI) sits at the center of modern video computing.

But once you move beyond demos and proof-of-concepts, the terrain looks very different. Real video is unpredictable. Environments keep shifting. Bandwidth is never guaranteed. And the systems must operate continuously without excuses. Few leaders understand this balance between intelligence and operational rigor better than Tuhin Bose, VP and CTO, Videonetics, who has spent over three decades building solutions for smart cities, transportation networks, and critical infrastructure.

Videonetics works in environments where video never stops, and reliability is nonnegotiable. The company's strength lies in blending deep AI with engineering discipline, especially across India's complex public spaces. In our discussion, he broke down how video analytics has changed, what challenges remain unsolved, and why responsible AI must be engineered from the ground up.

How video AI has transformed over two decades

The biggest transformation in video analytics has come from shifting rule-driven detection toward learning-driven intelligence. Early systems relied on rigid heuristics and fixed thresholds. They worked only when the environment behaved perfectly. As scenes grew busier and cameras became more widespread, these handcrafted methods ran out of room.

Modern AI, in contrast, learns directly from data. It interprets movement, behavior, and context in ways that were not possible earlier. Yet this evolution brings dependencies. Models need large, diverse datasets and far more computational complexity. In a country like India, where lighting, density, and camera quality change rapidly, these demands compound. For Videonetics, this shift meant moving from algorithm-level tuning to full-stack system design that supports continuous training and long-term reliability.

Why high-quality video data remains the toughest challenge