Principal Engineer: Annapurna Labs Training and Inference , Annapurna Labs, AWS Neuron SDK

At Amazon, Principal Engineers are both technical leaders and hands-on builders. As Amazon's most senior individual contributors, Principal Engineers work on our hardest problems, help establish technical standards, and drive the overall technical architecture, engineering practices, and methodologies for products that delight our customers. On the AWS Neuron SDK team within Annapurna Labs, that mandate is AI model performance on Amazon's custom silicon.

We are looking for a Principal Engineer to drive AI model performance across the lifecycle on AWS Trainium — pre-training, post-training and alignment, and inference serving — with deep hands-on expertise in one of these areas and the drive to influence across all three. Neuron is Amazon's software stack: the layer between customer ML frameworks and Trainium / Inferentia silicon. The responsibilities are broad: you will have the opportunity to drive deeply into model enablement across the path from a model's first pre-training step through aligned post-training to cost- and latency-optimized serving  will drive performance for a wide range of AI models, including the leading open-source models and direct collaborations with foundation-model developers, as well as supporting Amazon's own internal model development initiatives.

This is a deep hands-on builder role: you will write code, own the forward and backward path, and make binding technical decisions, while influencing direction across the broader Neuron organization, partner teams (EC2, EKS, VPC, Bedrock, Hyper Pod), and leading ML customers. Annapurna offers full-stack ownership — chip to server to network to compiler to cloud — so you see your work run at customer scale, something that doesn't exist at a standalone chip company or startup.

Over time, this person ends up owning the whole process end-to-end.

Key job responsibilities

As the Principal Engineer for Annapurna Labs Training and Inference, you will bring deep hands-on expertise in one of the following areas and influence across all three:

Pre-Training Performance Kernel optimization for ML training — forward and backward path — and drive performance out of it, with training collective algorithms (collectives, all-reduce, cross-accelerator coordination) as the primary source of gains.

Prove top-tier performance even at small scale, and clear the path so customers can land AI models on the chip out of the box. This is a performance role, not an Applied Science role — you partner with scientists but own making it fast, not making it correct.

Post-Training & Alignment Supervised fine-tuning, preference optimization, reinforcement learning, and alignment stack that turns pretrained base models into production-grade, aligned models on Trainium — the main lever on delivered model quality and the fastest-growing, least-covered workload class on Trainium today.

Deliver first-class RLHF/RLAIF and preference methods (PPO, GRPO, DPO/IPO/KTO, reward modeling) with training-plus-inference co-location and convergence safeguards (KL control, reward-scaling, policy-collapse detection), and native performance for the post-training ecosystem (TRL, veRL, NeMo-Aligner, Torch Titan/TorchAO, JAX).Inference / Serving Inference optimization strategy within the Neuron SDK — adapting inference algorithms to Trainium's unique architecture to unlock performance existing industry approaches leave on the table, and maximizing the chip's compute and memory bandwidth to drive cost and latency improvements for large-scale model serving.

Influence hardware-software co-design by partnering with Annapurna's chip architecture teams so future silicon incorporates lessons from production model serving.

Influence across all three

Serve as a technical leader on…