Principal Software Engineer - LLM Optimization

At JPMorganChase, we are building the infrastructure that powers the next generation of enterprise AI — and we need the best minds in LLM inference to help us do it. This is your opportunity to work at the intersection of cutting-edge machine learning and large-scale production systems, directly influencing how one of the world's largest financial institutions deploys and optimizes AI at scale. As a Principal Software Engineer at JPMorganChase within the AI/ML Data Platform team, you will serve as the firm's deepest technical voice on LLM inference performance — owning optimization strategy, benchmarking rigor, and efficiency at scale. You will work directly with senior engineering leadership to shape how our platform evolves, ensuring every model we serve is fast, cost-efficient, and production-ready. This is a high-visibility individual contributor role where your technical decisions will have direct, measurable impact on the firm's AI capabilities Job Responsibilities Own systematic benchmarking and performance characterization across all production LLM workloads. Establish reproducible baselines, catch regressions early, and quantify the impact of every configuration change before it touches production Design and execute quantization experiments — FP8, INT8/INT4 (GPTQ/AWQ), next-generation precision formats on current hardware — measuring accuracy delta, throughput improvement, memory reduction, and cost-per-token impact Drive speculative decoding strategy across the model portfolio: draft model, n-gram, and multi-token prediction approaches. Own acceptance rate measurement and per-workload configuration recommendations Build and maintain a GPU efficiency scorecard: utilization, memory headroom, cost per 1K tokens, and waste identified — giving leadership a data-driven view of platform efficiency at all times Benchmark our platform against external providers and published industry numbers — know what good looks like, and close the gap Lead inference engine upgrade evaluations: new scheduler architectures, async tensor parallelism, disaggregated prefill/decode, advanced speculative decoding — systematic validation before production promotion Collaborate with the EKS and disaggregated serving teams on KV-cache optimization, prefix caching strategies, and multi-node serving architecture Design and run GPU chaos engineering: induced failure scenarios, hardware diagnostic monitoring, detection and recovery measurement Architect and govern agentic AI-enabled engineering workflows (using enterprise-authorized tools within the work environment) to improve delivery speed, code quality, and operational outcomes at scale (e.g., AI-driven PR review assistance, test generation/maintenance, release readiness checks, incident triage and root-cause acceleration), while defining guardrails for validation, security, resiliency, and reuse across teams. Apply knowledge of tools within the Software Development Life Cycle toolchain, including enterprise-authorized AI-assisted development and automation capabilities, to improve the value realized by automation at scale. Required qualifications, capabilities, and skills Formal training or certification on software engineering concepts and 7+ years applied experience Deep, hands-on experience with LLM inference systems — vLLM, TensorRT-LLM, SGLang, LLM-D or equivalent production serving engines Strong grasp of GPU memory architecture: KV cache sizing and dynamics, memory-bandwidth vs compute bottlenecks, the practical implications of quantization at inference time Experience with quantization techniques and their real-world tradeoffs at scale Familiarity with speculative decoding and the variables that drive acceptance rates in production workloads Rigorous benchmarking instincts — GuideLLM, custom harnesses, or equivalent. Every claim has a number behind it Comfort operating in cloud GPU infrastructure at scale (AWS; EKS, managed inference services) Demonstrated awareness of