24 hours ago
San Francisco, CA, USA or San Jose, CA, USAEntry Level
Responsibilities
- Design, build, train, and evaluate AI/ML models using Python, TensorFlow, PyTorch, or JAX.
- Develop training pipelines with data loading, checkpointing, logging, validation, and experiment tracking.
- Work with CNNs, Transformers, LLMs, embeddings, and attention-based architectures.
- Debug convergence, overfitting, unstable loss, NaNs, tensor-shape errors, and GPU memory failures.
- Optimize GPU workloads involving CUDA execution, memory usage, batching, mixed precision, and kernel performance.
- Profile training and inference workloads to identify compute, memory, data-loading, and communication bottlenecks.
- Support custom operators, fused kernels, CUDA/Triton optimizations, and framework-level performance improvements.
- Build inference pipelines focused on latency, throughput, reliability, cost, and quality.
- Create evaluation pipelines for accuracy, robustness, hallucination risk, safety, latency, and regression testing.
- Read research papers, implement ideas, run experiments, and document findings, trade-offs, and limitations.
Requirements
- Strong programming experience in Python and understanding of data structures, algorithms, numerical programming, and clean code practices.
- Hands-on experience with TensorFlow, PyTorch, or JAX.
- Strong fundamentals in deep learning, backpropagation, gradient descent, optimizers, loss functions, regularization, CNNs, Transformers, attention mechanisms, embeddings, and model evaluation.
- Experience training and debugging models on CUDA-enabled GPUs.
- Understanding of GPU memory, CUDA kernels, threads, blocks, warps, memory bandwidth, kernel launch overhead, mixed precision training, and CUDA out-of-memory debugging.
- Comfort with Linux, Git, Docker, shell scripting, and experiment management.
- Preferred qualifications include CUDA C/C++ or Triton development, C++ for performance-critical AI systems, distributed training, GPU profiling, inference optimization, quantization, TensorRT, ONNX, XLA, model compilation, LLMs, RAG systems, agents, embeddings, multimodal AI, and AI evaluation frameworks.
- Ability to reason from first principles, solve ambiguous technical problems, implement research papers, profile before optimizing, and communicate technical findings clearly.
Benefits
- The role is on-site in San Francisco, United States.
About Accellor
Accellor is an IT services and consulting firm that designs and builds enterprise applications and customer experience solutions, with a focus on Salesforce, Microsoft Azure, mobile, and AI/chatbots. It provides systems integration, digital transformation, and product development for global enterprises across sectors, delivering solutions used in real operating environments. The company is privately held and headquartered in Fremont, California.