Frame Skipping to Cut Dhrishti Compute Costs
Reduced compute and infrastructure cost by skipping or reusing redundant frames.
Karan Mungra
Founding Engineer, profound.me
From their time as
SDE
Krutrim • Jan 2024 – Mar 2026
Overview
Karan identified that Dhrishti was recomputing results for frames that could be skipped or approximated without changing the analysis output. By caching and reusing values across similar frames, he reduced compute and cost by approximately 50 to 60%. He validated the change against historical data using the system's existing hard metrics.
The Story
While working on Dhrishti's GPU-backed camera-processing platform, Karan looked at where the system was doing redundant work. He found that certain frames could be skipped or replaced with approximate values from nearby frames without meaningfully affecting the analysis result.
He implemented frame skipping and reuse so the system avoided recomputing results it had already produced or could closely approximate. The change reduced both compute usage and infrastructure cost by roughly 50 to 60%, though that figure is his own estimate rather than a formally published measurement.
To check that analysis quality stayed acceptable, Karan tested the optimized system against historical data using the platform's existing hard metrics. The outputs matched what the system had produced before the change, giving him confidence that frame reuse did not degrade the analysis.
Ownership Snapshot
Broad role
Identified and shipped a compute optimization for Dhrishti's camera-processing pipeline.
Goal
Reduce GPU compute and infrastructure cost without degrading analysis quality.
Direct ownership
Owned the analysis of redundant frame computation, the implementation of frame skipping and reuse, and quality validation against historical data.
Team execution
Worked within the broader Dhrishti platform team operating custom models on GPU clusters.
Topics
- GPU Infrastructure
- Cost Optimization
- Video Processing
- Caching
Talk to Karan about this story
Send a message