Karan designs backend systems that retain work and recover predictably when components fail. In Dhrishti, he combined tenant isolation, ready backup nodes, lifecycle checks, and failure queues. In the training-data pipeline, he used event-driven stages, dedicated failure queues, versioned processing, and targeted reruns. Across both systems, he made recovery an explicit part of the architecture rather than an operational afterthought.

Karan builds the systems around AI products, from data preparation and GPU-backed serving to agent orchestration, tools, and memory. He co-built a multi-terabyte training-data pipeline and designed distributed camera processing across GPU clusters. He also built a C++ memory service, worked on tool integration for Krutrim's assistant, and built the multi-agent orchestration layer that decomposed complex queries across six to eight specialized agents. His contribution sits at the boundary between backend engineering, model-serving needs, and production reliability.

Karan creates concrete checks for AI components that need measurable behavior. For the assistant memory system, he built an evaluation pipeline around production queries and expected memory behavior. For the voice-to-voice product at profound.me, he built a dedicated evaluation pipeline, used it to surface ASR misspelling issues, and drove iterative fine-tuning that resolved user-reported cases and approximately 95% of evaluation-identified cases by his recollection. In agent tooling and orchestration, he tracked tool-call success and fixed a recursive scheduling bug that caused runaway agent behavior. For the Dhrishti frame-skipping optimization, he validated quality by testing against historical data using the platform's hard metrics, confirming that approximate frame values produced the same outputs as full recomputation.