Not one role — 232 real RE job descriptions cluster into eight archetypes, and each
expects a different Pareto trade-off and skill set. Click an archetype to see what it actually asks for.
Drill into an archetype
Inference, performance & kernels 36 Post-training & alignment 35 Evals, safety & interpretability 31 Agents & environments 22 Applied research & product bridging 22 Data & synthetic-data engineering 19 Distributed training systems 19 Multimodal & embodied AI 14 Research platforms & reproducibility 13 Model & algorithm research 11 Scientific AI 10
Pick an archetype above to see its must-have vs nice-to-have capabilities, methods, systems, and expected artifacts.
Inference, performance & kernels · 36 JDs
Dominant Pareto axes: Latency 94% · Capability 92% · Cost/compute 83% · Reliability 64%
Must-have capabilities (share of JDs) GPU/CUDA/kernels 61% Python 42% Distributed training 33% Inference/serving 28% C++/Rust 14% Research/papers 11% LLM/transformers 11% Agents/tools 11%
Nice-to-have capabilities GPU/CUDA/kernels 56% Inference/serving 42% LLM/transformers 33% Distributed training 22% Agents/tools 14% PyTorch 11%
Top methods benchmarking Kernel optimization Hardware-software co-design speculative decoding Performance profiling performance profiling quantization Kernel profiling
Top systems PyTorch GPU clusters MLIR CUDA Cerebras WSE Hugging Face Transformers Nsight NVIDIA GPU architectures
Expected artifacts (what you'd produce) Benchmark reports Performance models Profiling reports performance reports performance benchmarks
Post-training & alignment · 35 JDs
Dominant Pareto axes: Capability 100% · Reliability 100% · Data/coverage 83% · Safety 51%
Must-have capabilities (share of JDs) Python 57% Research/papers 43% Post-training/finetune 37% PyTorch 26% Distributed training 26% RL 23% LLM/transformers 23% Evals/benchmarks 17%
Nice-to-have capabilities RL 57% Post-training/finetune 43% LLM/transformers 40% Distributed systems 29% Evals/benchmarks 29% Agents/tools 29%
Top methods RLHF RLHF/RLAIF SFT Synthetic data generation reinforcement learning RL/RLHF/RLAIF Reinforcement learning evaluation design
Top systems PyTorch Post-training stack agent harnesses RL training infrastructure evaluation frameworks Training pipelines Agent harnesses Experiment tracking
Expected artifacts (what you'd produce) evaluation suites evaluation frameworks Optimized post-training recipes Alignment experiments Synthetic data and data mixtures
Evals, safety & interpretability · 31 JDs
Dominant Pareto axes: Reliability 100% · Capability 87% · Safety 77% · Data/coverage 71%
Must-have capabilities (share of JDs) Research/papers 68% Python 45% Evals/benchmarks 32% Distributed training 23% Distributed systems 19% Data/pipelines 19% PyTorch 13% RL 10%
Nice-to-have capabilities Research/papers 52% Evals/benchmarks 45% Agents/tools 35% LLM/transformers 26% Distributed training 19% Data/pipelines 19%
Top methods Statistical analysis Mechanistic interpretability Adversarial testing anomaly detection environment design Experimental design Evaluation framework design Benchmarking
Top systems PyTorch JAX CUDA Sparse autoencoders Evaluation pipelines Evaluation infrastructure Data pipelines Anthropic Model APIs
Expected artifacts (what you'd produce) Research papers detection classifiers adversarial test sets safety evaluation reports prototyped safeguards
Agents & environments · 22 JDs
Dominant Pareto axes: Capability 100% · Reliability 100% · Safety 68% · Data/coverage 64%
Must-have capabilities (share of JDs) Python 59% RL 36% PyTorch 32% Agents/tools 27% Research/papers 27% Distributed systems 23% Post-training/finetune 23% JAX 14%
Nice-to-have capabilities Agents/tools 50% RL 32% LLM/transformers 32% Research/papers 32% Distributed systems 23% Evals/benchmarks 18%
Top methods Reinforcement learning Sim-to-real transfer RL Experimental design Distributed systems engineering simulation Data engineering Policy evaluation
Top systems PyTorch Kubernetes Distributed training systems Simulation environments Agent harness Python C++ Isaac Sim or MuJoCo
Expected artifacts (what you'd produce) Novel training environments Evaluation frameworks RL policies for manipulation and locomot Data pipelines for training and evaluati Deployment-ready robot skills
Applied research & product bridging · 22 JDs
Dominant Pareto axes: Capability 100% · Reliability 95% · Data/coverage 73% · Latency 64%
Must-have capabilities (share of JDs) Python 55% LLM/transformers 27% Agents/tools 23% Evals/benchmarks 18% Research/papers 18% C++/Rust 14% Distributed training 14% PyTorch 14%
Nice-to-have capabilities Agents/tools 50% LLM/transformers 23% C++/Rust 14% PyTorch 14% Evals/benchmarks 9% Inference/serving 9%
Top methods Prompt engineering Sensor fusion Containerization Model selection Tool design Performance optimization Computer vision Machine learning
Top systems Embedded Linux PyTorch Lattice OS Autonomous maritime vessels Radar/sonar/LIDAR/camera sensor suites Simulation frameworks Autonomous vehicle stack ML infrastructure
Expected artifacts (what you'd produce) Perception and planning software Simulation and test infrastructure Reusable code across product lines Data-analysis tooling Autonomous maritime capability updates
Data & synthetic-data engineering · 19 JDs
Dominant Pareto axes: Capability 100% · Data/coverage 100% · Reliability 74% · Cost/compute 37%
Must-have capabilities (share of JDs) Data/pipelines 58% Python 53% Research/papers 32% Distributed systems 26% Distributed training 26% LLM/transformers 21% PyTorch 16% Evals/benchmarks 11%
Nice-to-have capabilities Data/pipelines 37% Distributed training 21% Research/papers 16% LLM/transformers 16% Evals/benchmarks 11% Python 11%
Top methods Synthetic data generation data ablations data curation synthetic data generation Generative ML World modeling Sensor modeling Evaluation/optimization
Top systems Spark Kubernetes Ray Apache Beam multimodal data pipelines Sensor simulation stack Rendering systems Physics-based modeling tools
Expected artifacts (what you'd produce) Synthetic data pipelines high-quality multimodal datasets Sensor-simulation ML models Production simulator updates Evaluation and optimization reports
Distributed training systems · 19 JDs
Dominant Pareto axes: Reliability 95% · Capability 79% · Cost/compute 79% · Latency 58%
Must-have capabilities (share of JDs) Distributed training 84% Python 58% Distributed systems 47% PyTorch 32% GPU/CUDA/kernels 26% JAX 21% C++/Rust 16% RL 11%
Nice-to-have capabilities Distributed systems 63% GPU/CUDA/kernels 47% Distributed training 32% Research/papers 26% PyTorch 26% RL 21%
Top methods performance profiling performance optimization Performance profiling distributed systems debugging data pipeline engineering profiling API design distributed training
Top systems Kubernetes GPU clusters PyTorch JAX FSDP Megatron-LM distributed training systems NCCL
Expected artifacts (what you'd produce) distributed training systems Benchmark suites Performance profiling reports production-ready multimodal models scalable training components
Multimodal & embodied AI · 14 JDs
Dominant Pareto axes: Capability 100% · Reliability 100% · Data/coverage 93% · Safety 64%
Must-have capabilities (share of JDs) Python 93% PyTorch 57% JAX 43% Evals/benchmarks 29% Research/papers 29% Distributed training 14% Data/pipelines 14% LLM/transformers 7%
Nice-to-have capabilities Research/papers 57% Distributed training 14% GPU/CUDA/kernels 14% Agents/tools 14% RL 14% LLM/transformers 14%
Top methods 3D computer vision VLMs Evaluation design Multimodal fusion and alignment Representation learning Structured prediction Long-horizon reasoning modeling World modeling
Top systems PyTorch JAX Spark Ray ROS/ROS2 Multimodal foundation-model architecture GPU clusters Robot autonomy simulation
Expected artifacts (what you'd produce) Evaluation benchmarks Model architectures for perception/reaso Trained multimodal foundation models Experimentation frameworks Evaluation metrics
Research platforms & reproducibility · 13 JDs
Dominant Pareto axes: Reliability 100% · Maintainability 100% · Capability 69% · Cost/compute 62%
Must-have capabilities (share of JDs) Python 77% Distributed training 54% Distributed systems 46% Agents/tools 31% GPU/CUDA/kernels 23% Inference/serving 8% RL 8% PyTorch 8%
Nice-to-have capabilities Distributed systems 46% Agents/tools 46% GPU/CUDA/kernels 31% Research/papers 23% Distributed training 23% LLM/transformers 23%
Top methods Rapid prototyping Data pipeline engineering Performance profiling CI/CD optimization Agentic framework integration MCP server design Developer-experience instrumentation Systems and architecture design
Top systems Kubernetes PyTorch JAX TensorFlow DeepSpeed SLURM Distributed training systems Bazel
Expected artifacts (what you'd produce) MCP servers Autonomous internal agents CI/CD and build-tooling improvements Internal developer platform components Developer-experience dashboards
Model & algorithm research · 11 JDs
Dominant Pareto axes: Capability 100% · Reliability 82% · Data/coverage 73% · Cost/compute 55%
Must-have capabilities (share of JDs) Python 55% Research/papers 36% Evals/benchmarks 18% PyTorch 18% JAX 18% Distributed systems 9% Data/pipelines 9% Post-training/finetune 9%
Nice-to-have capabilities LLM/transformers 55% Research/papers 55% RL 9% Agents/tools 9% Distributed systems 9% PyTorch 9%
Top methods ablation studies large-scale experimentation architecture design Architecture design Training-system implementation Inference-time conditioning Retrieval augmentation Performance profiling
Top systems JAX TensorFlow PyTorch Data preprocessing pipelines Training and inference systems Embedding model stack PyTorch training pipeline Dataset generation tooling
Expected artifacts (what you'd produce) architecture prototypes Infrastructure for next-generation LLM r Scalable ML pipelines Open-source tools and model releases Research methodology for evaluating flex
Scientific AI · 10 JDs
Dominant Pareto axes: Capability 100% · Reliability 90% · Data/coverage 90% · Cost/compute 50%
Must-have capabilities (share of JDs) Python 60% Research/papers 50% JAX 40% PyTorch 40% Distributed training 30% LLM/transformers 20% Distributed systems 10% Agents/tools 10%
Nice-to-have capabilities Research/papers 60% Distributed systems 30% Agents/tools 30% Distributed training 30% GPU/CUDA/kernels 20% JAX 10%
Top methods JAX PyTorch Distributed training Supervised learning Probabilistic modeling Graph-based learning Optimization Learned simulation
Top systems JAX PyTorch XLA Triton CUDA Pallas GCP Kubernetes
Expected artifacts (what you'd produce) Scalable AI model prototypes Optimized data/training pipelines Research technical reports Software prototypes for real-world susta Optimized training pipelines