Principal Machine Learning Engineer, Conversational AI Modeling and Learning

Alexa AI is building the next generation of Alexa+, Amazon"s LLM-powered conversational assistant, and its future is agentic: LLM systems that reason and act over dozens of chained inferences, coupled to real environments where their actions persist. Making these agents smarter, faster, and cheaper is as much a systems problem as a modeling problem - agent performance depends on the model, the harness, the evaluation infrastructure, and the serving stack co-designed together.

We are looking for a Principal Engineer to lead the engineering of this agentic platform. You will own the architecture that turns research into production capability: large-scale agentic evaluation infrastructure (sandboxed, reproducible, statistically trustworthy at high concurrency), reinforcement learning training systems for long-horizon multi-turn trajectories, self-learning pipelines that convert production experience into permanent model and system improvements, and the serving architecture for latency-sensitive agentic inference. You will partner closely with scientists and work backwards from committed product launches, setting the technical bar for a platform that serves every Alexa agent rather than one product at a time.

The charter is the full lifecycle of a production agent: how it is measured, how it is trained, how it learns, and how it is served. You will build the harnesses and sandboxed worlds agents act in, the evaluation systems that make their quality provable rather than asserted, the RL infrastructure that trains models on the same tasks they are measured on, and the self-improvement loop that turns every production interaction into a permanently smarter system - agents that ship better than they launched, week over week. Few places let one engineer shape the entire loop from a customer"s spoken request to a model that learned from it; this role owns that loop at Alexa scale.

Key job responsibilities

Define and drive the engineering roadmap and architecture for the agentic AI platform: evaluation, training, self-learning, and serving for LLM-based agents in production

Architect large-scale agentic evaluation infrastructure: isolated sandboxed execution, recreatable environments, verifiable scoring, and reproducibility at hundreds of concurrent trials, so model decisions rest on trustworthy numbers

Build and scale RL and post-training systems for agentic workloads: 256K+ token contexts, multi-turn trajectory training, train/inference engine consistency, and reward attribution across long sessions

Design the serving and inference architecture for agentic traffic (long sessions, output-generation-bound workloads, KV-cache-centric optimization), co-designing with inference-infrastructure partner teams

Set the technical bar across the organization: raise engineering standards through design reviews, operational excellence, and deep dives on the hardest cross-system problems

Translate ambiguous product and science requirements into platform interfaces partner teams can build on; influence senior leadership on build-vs-adopt and ownership decisions

Mentor and grow senior and principal-track engineers across multiple teams

A day in the life

You might spend the morning in a design review for the next generation of the evaluation platform"s execution layer, midday debugging why a 100K-token training session diverges between the rollout engine and the learner, and the afternoon with the serving team deciding which KV-cache optimizations justify architectural investment before a product launch. You work daily with applied scientists and other engineers, and your systems are the reason their results are trustworthy and shippable.

About the team

Our organization owns the applied science and platform engineering for Alexa"s agentic experiences. We operate at the intersection of large language models, reinforcement learning with verifiable rewards, agentic architectures, and large-scale distributed systems, serving customers across dozens of languages and device types. Our platform provides the shared evaluation, training, self-learning, and serving foundation for Alexa"s flagship agent programs and the broader agent portfolio behind them.