TBC is building next-generation AI systems at the intersection of biological computing, generative models, and large-scale AI infrastructure. As we scale our world-model and neural-optimizer efforts, we are looking for an optimization-focused Research Scientist / ML Engineer to improve the efficiency, latency, throughput, and deployability of large models.
This role is focused on making frontier models run faster, cheaper, and more reliably — especially LLMs, diffusion models, video generation models, and world-model systems. You will work across inference optimization, training efficiency, model compression, memory management, and GPU-level performance to help turn research systems into scalable, customer-ready products.
WHAT YOU’LL WORK ON
- Optimize inference for LLMs, diffusion models, video models, and world-model systems
- Improve serving efficiency through techniques such as KV caching, batching, quantization, distillation, speculative decoding, and memory optimization
- Build and optimize high-throughput inference pipelines for large models running on GPU clusters
- Profile model performance across latency, throughput, memory usage, GPU utilization, and cost
- Implement custom kernels or low-level optimizations using Triton, CUDA, PyTorch, or related systems
- Improve training and fine-tuning efficiency for large generative models, including distributed training, checkpointing, parallelism, and data loading
- Work with research teams to identify bottlenecks in model architecture, inference paths, and deployment workflows
- Translate model performance improvements into clear customer-facing benchmarks and technical proof points
- Evaluate trade-offs across model quality, latency, cost, memory, and deployability
WHAT WE’RE LOOKING FOR
- Strong background in machine learning systems, model optimization, or high-performance AI infrastructure
- Hands-on experience optimizing LLMs, diffusion models, video generation models, or other large generative systems
- Experience with one or more of:
- Inference optimization
- KV caching / attention optimization
- Triton or CUDA kernel development
- Quantization, pruning, distillation, or model compression
- Distributed training / fine-tuning efficiency
- GPU profiling and performance debugging
- Strong PyTorch experience and comfort working close to the model/runtime boundary
- Ability to reason about trade-offs between quality, latency, throughput, memory, and cost
- Comfortable working across research code, production systems, and benchmarking infrastructure
- Excited to work in an ambiguous, early-stage environment where optimization work directly shapes product feasibility
WHAT SUCCESS LOOKS LIKE
- Large models run faster, cheaper, and more reliably across TBC’s core workloads