Surya Atmuri

Key Achievements

• Designing a RAG pipeline on LLM-guided evolutionary PyTorch architecture search on CIFAR-10, using 2 FAISS vector indices with a cross-encoder reranker to retrieve prior experiments and documentation for faster NAS convergence
• Migrated the Python inference server to utilize vLLM’s continuous batching and PagedAttention, accelerating throughput by 3.6x and achieving >95% KV cache utilization to serve Llama-3.3-70B models within 80GB VRAM limits
• Engineered a distributed fitness inheritance system with self-correcting LLM infrastructure on a SLURM cluster to reuse CNN checkpoints and autonomously resolve runtime errors, cutting 48+ redundant GPU evaluations per run