shekhar parmar

I’m a systems- and backend-oriented developer with a strong focus on production ML inference pipelines, performance engineering, and distributed state management. My work centers on building efficient, failure-tolerant systems for high-throughput workloads—especially multi-modal embedding pipelines and vector-search powered experiences.

In a production semantic deduplication system, I designed and operated a near-duplicate detection pipeline processing 4M+ messages for 10,000 concurrent users, delivering 35% spam reduction and 20% engagement improvement after deployment. I implemented adaptive CTPH chunking, integrated text (MPNet 768D) and image (DINOv2 768D) embeddings via ONNX Runtime with CUDA, and used Qdrant plus Redis configuration to enable fast similarity search—while engineering recovery, incremental index updates, and rate-aware synchronization to handle spikes. I also build high-traffic automation suites (C++/Node.js) with resilient, rate-limit aware event pipelines and consistent cross-server state.