Shivam Akhauri

I build the infrastructure that makes AI agents actually work in production. Currently at SolarWinds, I fine-tuned a Qwen3 14B reward model with PPO-based RLHF (KL-regularized, trained with PyTorch FSDP on A100 GPUs) for agentic tool routing -- improving tool selection from 72% to 89% and cutting hallucinated tool calls by 40%. I also designed SWAG, a federated agentic platform where 15 domain teams ship AI capabilities via MCP servers, composable agents, and skills, and architected a hybrid LLM serving layer (vLLM + Triton) running at p99 <200ms and 500 RPS. Before that at Ocrolus, I trained a custom early-fusion multimodal Transformer (LayoutLMv3-style with Longformer sparse attention, context extended from 512 to 4096 tokens) processing 50M+ financial documents at 97% macro-F1, and designed GPU serving infrastructure on EKS that saved $800K/year. My research spans agent safety and reliability, with 3 papers currently under review at NeurIPS 2026 (decision-accountability certificates for tool-use agents, drift bounds for self-modifying agents, and agent anomaly-detection benchmarking), building on prior work in domain adaptation for autonomous driving (IROS'20, PerCom CARD'25, 60+ citations). What I work on day-to-day: - Agentic AI systems and multi-agent orchestration (plan-based + autonomous) - RLHF, reward modeling (Bradley-Terry), and iterative alignment (PPO, KL control) - LLM inference optimization (vLLM, PagedAttention, continuous batching) - Multimodal Transformers and sparse attention architectures - GPU serving infrastructure at scale (Kubernetes, KEDA, Triton) Currently exploring: speculative decoding for structured tool-call outputs, prefill-decode disaggregation, and reward model drift detection. Open to Senior/Staff Machine Learning Engineer / Software Engineer ML roles focused on agentic systems, Large Language Model (LLM) infrastructure,