Forward Deployed Engineer

Who you are

• Experience in professional, non-internship software development
• Experience programming with at least one modern language such as Java, C++, or C# including object-oriented design
• 3+ years of non-internship professional software development experience
• 2+ years of experience working in an Agile, startup, or rapidly scaling environment
• Hands-on experience with generative AI technology, such as LLMs, vision-language models, or speech models
• Experience with AI tools: eval frameworks, agent tooling, RAG pipelines, and prompt engineering
• Experience communicating technical concepts to both technical and non-technical audiences
• Bachelor's degree in computer science or equivalent
• Experience with full software development life cycle, including coding standards, code reviews, source control management, build processes, testing, and operations
• Experience with large-scale distributed training of LLMs on multi-node accelerator clusters, using frameworks such as PyTorch FSDP2, TorchTitan, Megatron-LM/Megatron Core, NeMo, DeepSpeed, or NeuronX Distributed
• Experience with LLM post-training, including SFT, DPO, RLHF, and RL with verifiable rewards or multi-turn agentic RL, using frameworks such as Miles, slime, verl, NeMo-RL, OpenRLHF, SkyRL, or TRL
• Experience optimizing model training or inference for latency, throughput, or cost, using serving engines such as SGLang, vLLM, or TensorRT-LLM, on AWS Trainium or Inferentia (Neuron SDK), on AWS services such as SageMaker AI, SageMaker HyperPod, or EKS, or by writing custom kernels (CUDA, Triton, NKI)
• Experience deploying and operating ML models in customer production environments, working with teams that build and scale end-to-end production systems

What the job involves

• As a Forward Deployed Engineer II (FDE II) on the Custom Model and Optimization team, you embed directly with strategic customers to build, optimize, and run custom generative AI models in production alongside their engineering teams
• You own complete model workstreams end-to-end from defining the evaluation bar and preparing data, through training and inference optimization, to production deployment
• FDE IIs on this team works at the intersection of models and systems
• They customize state-of-the-art open-weight and foundation models for new domains, languages, and modalities; migrate and optimize training and inference workloads on AWS Trainium and operate those models as production services
• They are autonomous contributors who take on projects where the business goal is clear, but the path is not
• They make design trade-offs with appropriate guidance, limit short-term workarounds, coach others on producing high-quality code, and train new team members
• Embed within customer ML and engineering teams: Understand the customer's business, data, model stack, and operational constraints. Translate ambiguous goals into a measurable evaluation bar and a technical plan for your workstream such as model selection, customization approach, and serving architecture. Requires up to 25% travel
• Customize and post-train models. Adapt open-weight and foundation models to customer domains through continued pre-training, supervised fine-tuning, preference optimization, and reinforcement learning (including RLHF, RL with verifiable rewards, and multi-turn agentic RL). Build the data pipelines, synthetic data generation, and evaluation harnesses that prove a customized model out performs the baseline on the customer's own quality, latency, and cost bar. Apply these techniques to high-value patterns such as coding agents, domain foundation models, and multi-adapter serving
• Build large-scale training and inference systems. Design and implement distributed training and RL post-training pipelines using frameworks such as PyTorch FSDP2, Megatron-LM, Miles, and verl. Build RL interaction harnesses using NeMo Gym or OpenEnv to connect agent workflows, tools, environments, and reward functions to training. Develop production inference stacks with engines such as SGLang and vLLM, and optimize throughput, latency, memory footprint, and cost per token through quantization, parallelism, speculative decoding, and batching
• Optimize models on AWS chips. Migrate and tune training and inference workloads for AWS Trainium using the AWS Neuron SDK. Profile performance, close accuracy-parity gaps against GPU baselines, and develop custom kernels where needed. Benchmark rigorously and turn results into decisions customers can act on
• Accelerate impact beyond the engagement. Codify reusable fine-tuning recipes, evaluation frameworks, benchmarking tools, and optimization patterns that accelerate future deployments. Feed field learnings back to AWS service teams, including Amazon Bedrock, Amazon SageMaker AI, and AWS Neuron. Identify opportunities to expand model customization and optimization within your customer engagement