GPU Software Engineer; CUDA

Bright Vision Technologies is a forward-thinking software development company dedicated to building innovative solutions that help businesses automate and optimize their operations. We leverage cutting‑edge technologies to create scalable, secure, and user‑friendly applications.

We’re looking for a skilled GPU Software Engineer (CUDA) to join our dynamic team and contribute to our mission of transforming business processes through technology.

Job Title: GPU Software Engineer (CUDA)

Location: 100% Remote (Continental United States)

Position Type: In‑house Bright Vision Technologies SOW engagement (no third‑party client or vendor)

Experience: 6+ years

Salary: 100K – 150K

Sponsorship: No new H1B sponsorship available. H1B transfers welcomed for qualified candidates.

Employment Type: Full‑time, direct W2 with Bright Vision Technologies (no C2C, no 1099, no third‑party)

Engagement: Long‑term, multi‑year, aligned to the Bright Vision SOW delivery roadmap

Compensation: Competitive base salary commensurate with experience, plus benefits.

This is a 100% remote, full‑time, direct W2 position with Bright Vision Technologies.

This role is part of Bright Vision Technologies’ in‑house Statement of Work (SOW) engagement. The client, end customer, and employer for this position is Bright Vision Technologies — there is no third‑party client, vendor, or implementation partner involved.

We do not engage in C2C, 1099, or third‑party arrangements for this role.

However, candidates who are currently on a valid H1B visa and require a transfer are welcome to apply. We will support H1B transfers for qualified candidates.

For every role, a technical coding assessment is mandatory. Please apply only if you are confident in your technical abilities and hands‑on experience.

We are seeking a GPU Software Engineer (CUDA) with deep expertise in CUDA programming, GPU architecture, and high‑performance computing to design and optimize compute‑intensive workloads on modern accelerator hardware. This role focuses on extracting maximum performance from GPU platforms for AI training, inference, scientific computing, and high‑throughput data processing workloads. The ideal candidate combines low‑level systems mastery with strong software engineering practices, and has a track record of delivering measurable performance improvements on production GPU systems.

In this role you will work closely with cross‑functional partners — product, design, engineering, operations, and business stakeholders — to translate ambiguous requirements into well‑engineered solutions, and will be expected to raise the bar through code review, design review, and mentorship of more junior engineers. The successful candidate brings strong engineering discipline, a clear communication style, and a track record of shipping meaningful work that holds up well in production.

• Design and implement high‑performance CUDA kernels for compute‑intensive workloads across AI and HPC use cases.
• Profile and optimize GPU code using tools such as Nsight Systems, Nsight Compute, and CUDA profilers.
• Tune memory access patterns, occupancy, register usage, and shared memory utilization for peak performance.
• Develop highly optimized libraries for linear algebra, attention, and other ML primitives.
• Optimize multi‑GPU and multi‑node training using NCCL, RDMA, and high‑performance networking.
• Implement custom operators and fused kernels in PyTorch, JAX, or Triton.
• Collaborate with ML engineers to identify performance bottlenecks in training and inference pipelines.
• Develop benchmarks and regression tests to safeguard performance over time.
• Evaluate new GPU architectures and feature sets, and advise on adoption strategy.
• Contribute to compiler‑level optimizations for tensor programs where appropriate, working at the boundary between ML frameworks and underlying accelerator codegen to unlock performance not reachable through framework‑level tuning alone.
• Optimize memory hierarchy usage across HBM, L2, shared memory, and registers.
• Implement mixed‑precision and quantized compute paths that maximize accelerator throughput while preserving numerical…