Atla Smaran Reddy.

Designed and deployed a production-grade adaptive RAG pipeline with parallel retrieval and confidence-based fast-path execution that bypasses reranking on high-confidence queries. Improved reranking efficiency via batched LLM cross-encoder scoring and engineered type-aware failure classification with a self-correcting retry mechanism.

Built an incident diagnosis system that retrieves and ranks past production failures by semantic + structured similarity, surfacing relevant root causes and resolutions for on-call engineers. Designed a pipeline to convert unstructured incident reports and discussions into structured signals.

Built a hybrid CNN–ViT–GNN object detection system with query-based decoding, enabling relational reasoning and context-aware multi-object prediction. Implemented dynamic graph-based feature learning, custom detection loss, and scalable PyTorch training pipeline with GPU support and checkpointing.