Spurthi Sandiri

I solve complex, ambiguous problems at the intersection of AI models, hardware, and distributed systems — and ship solutions that serve millions of customers. As a Senior SDE at Amazon, I lead inference optimizations for the Amazon Nova family of models — spanning GPU and custom silicon, from single-node to multi-node frontier scale. LLM inference is where research and production speak different languages. A technique that shows gains in a paper often falls apart against real hardware constraints, feature cross-compatibility (multimodal, customization, constrained decoding), quantization trade-offs, and the tension of keeping accelerators maximally utilized while serving low-latency responses. Add external dependencies with uncertain timelines and multiple teams with conflicting hypotheses — and the right approach often doesn't exist yet. That's where I thrive: → Led speculative decoding across Amazon Nova models — significant throughput improvements with multimodal, constrained decoding, and LoRA cross-compatibility → Unblocked a frontier-scale model launch by debugging accuracy regressions and delivering major throughput gains under multi-node constraints → Implemented speculative decoding for LoRA, enabling low-latency model customization for a critical product launch → Root-caused a critical latency regression, then delivered substantial performance improvements → Wrote custom CUDA kernels for performance-critical inference paths → Filed 3 patents in inference optimization How I work: → End-to-end ownership from ambiguous requirements through production deployment → Bias for action — I make decisions with incomplete data, move fast, and course-correct → Cross-team alignment across model training, evals, runtime orchestration, and external partners → Force multiplier through mentoring, knowledge sharing, and scalable frameworks → T