AI Performance Engineer

Bright Vision Technologies•South Plainfield, NJ

20h•$100,000 - $150,000•Remote

About The Position

Bright Vision Technologies is seeking an AI Performance Optimization Engineer to improve throughput, latency, and cost across training and inference workloads for large neural network systems. The role spans low-level kernel optimization through distributed-system tuning, drawing on GPU architecture, model parallelism, memory management, and compiler-level optimization. The engineer will use rigorous instrumentation and measurement to make data-driven decisions and deliver improvements to production AI workloads. The position works with product, design, engineering, operations, and business stakeholders to turn ambiguous requirements into well-engineered solutions. It also involves raising engineering standards through code and design reviews and mentoring junior engineers. This is a full-time, direct W2 role with Bright Vision Technologies, performed remotely within the continental United States as part of the company’s in-house, long-term SOW engagement. The role does not involve a third-party client, vendor, or implementation partner. Salary is listed as $100,000–$150,000 per year, with benefits.

Requirements

Bachelor’s or Master’s degree in Computer Science, Computer Engineering, or a related field.
Six or more years of experience in performance engineering, ML systems, or HPC.
Strong proficiency in Python and C++.
Hands-on experience optimizing deep learning workloads on modern GPUs.
Deep understanding of distributed training and inference techniques.
Experience with profiling tools across CPU, GPU, and distributed systems.
Familiarity with model compression techniques and their accuracy implications.
Strong grasp of memory hierarchies, communication primitives, and parallelism strategies.
Excellent measurement, debugging, and analytical reasoning skills.
Strong communication and collaboration skills.
Candidates must be willing to work directly as a full-time W2 employee of Bright Vision Technologies.
A technical coding assessment is mandatory.
No new H1B sponsorship is available; qualified candidates currently on a valid H1B visa who require a transfer are welcome to apply.

Nice To Haves

Experience optimizing LLM inference at production scale.
Contributions to vLLM, TensorRT-LLM, DeepSpeed, or similar projects.
Familiarity with custom kernel authoring in Triton or CUTLASS.
Experience with FinOps for AI workloads.
Publications or talks on AI systems performance.

Responsibilities

Profile and optimize end-to-end AI training and inference pipelines for throughput, latency, and cost.
Identify and eliminate bottlenecks in data loading, model computation, communication, and memory.
Implement and tune quantization, sparsity, and pruning strategies to reduce model footprint and accelerate inference.
Optimize distributed training using tensor parallelism, pipeline parallelism, FSDP, and ZeRO-style sharding.
Tune attention implementations, including FlashAttention and paged attention.
Implement KV cache optimization, continuous batching, and speculative decoding for LLM serving.
Drive compiler-level optimizations using Triton, XLA, TorchInductor, or TVM, and work with the broader ML framework community on measurable end-to-end performance improvements.
Optimize data pipelines, sharding strategies, and storage access patterns for high-throughput training.
Build and maintain benchmark suites and regression frameworks across workloads.
Collaborate with ML and platform engineering teams to embed best practices in standard pipelines.
Improve cost efficiency through model architecture, hardware selection, and scheduling strategies.
Evaluate new hardware and software offerings and advise on adoption.
Document performance-tuning playbooks and share findings across engineering teams.
Stay current with AI systems research and apply relevant advances to production.
Participate in code reviews and design reviews, and mentor more junior engineers.