The Role
Optimize large-model inference at scale: improve throughput, reduce latency and cost per token, build benchmarking harnesses, tune parallelism and quantization strategies, implement load-balancing in routing, and evaluate custom kernels and emerging inference hardware.
Summary Generated by Built In
About Dragonfly
Dragonfly is a global investment firm focused on the future of money, markets, and intelligence. With origins in crypto and over $3.5 billion in assets under management, we invest in generational companies and protocols across the frontier of finance and technology. We're a team of seasoned VCs, technologists, and markets experts with deep roots in New York, Silicon Valley, and Asia. We've published some of the industry's most widely read research on stablecoins, market structure, decentralized finance, networks, protocol security, and increasingly, the intersection of finance and AI.
This is an application to join our talent network. This is not a listing for an internal role at Dragonfly.
We're actively sourcing for a Senior Inference Optimization Engineer for one of our portfolio companies building privacy-first consumer AI infrastructure. You'll be on the bleeding edge of LLM inference performance, pushing throughput, driving down latency, and optimizing cost per token at significant scale.
Location: Remote, USA (open to excellent candidates outside the USA)
What We’re Looking For:
- 5+ years in performance optimization or HPC with deep GPU architecture and parallel programming knowledge
- Hands-on experience with at least one production LLM inference engine (vLLM, SGLang) running at high volume
- Demonstrated experience with LLM inference optimization: continuous batching, PagedAttention, KV cache management, speculative decoding, quantization, CUDA graphs, torch.compile
- Experience with distributed inference strategies: tensor parallelism, pipeline parallelism, MoE parallelism in multi-GPU and multi-node environments
- GPU profiling fluency: Nsight Systems, Nsight Compute, PyTorch Profiler
- Proficiency in Python, Rust, or Go. C++/CUDA a strong plus
- Bonus: custom Triton kernels, diffusion/image model inference optimization, open-source inference framework contributions
About the role:
- Stand up and optimize GPU infrastructure including B300 nodes in owned data centers
- Drive down TTFT and TPOT, push throughput, and improve cost per token for LLM inference workloads
- Build reproducible benchmarking harnesses across inference engines to identify optimal engine, quantization scheme, and parallelism strategy per workload and GPU SKU
- Optimize multivariate inference load-balancing algorithms within the inference routing system
- Evaluate emerging inference optimization techniques including custom CUDA/Triton kernels, novel attention variants, new quantization schemes, and compilation stack improvements
- Evaluate emerging inference hardware (FPGAs, ASICs, custom silicon) for viability in the stack
Even if you don't match every point above, we encourage you to apply. There may be other opportunities that fit your skill set.
Process:
- We'll review your application and assess fit for this role.
- If there's a match, we'll facilitate a warm introduction to the team.
- If the timing isn't right, we'll keep you in mind for future opportunities across the portfolio.
Compensation
- Early Career: $110,000–$150,000
- Mid: $150,000–$200,000
- Senior: $180,000–$250,000
- Staff/Leadership: $230,000–$330,000
The compensation range reflects variation by seniority, location, and hiring company. Final compensation is confirmed with the specific hiring company.
Submit your information below, and we’ll reach out if there’s a potential fit.
Skills Required
- 5+ years in performance optimization or HPC with deep GPU architecture and parallel programming knowledge
- Hands-on experience with a production LLM inference engine (vLLM, SGLang) running at high volume
- Experience with LLM inference optimization: continuous batching, PagedAttention, KV cache management, speculative decoding, quantization, CUDA Graphs, torch.compile
- Experience with distributed inference strategies: tensor parallelism, pipeline parallelism, MoE parallelism in multi-GPU and multi-node environments
- GPU profiling fluency: Nsight Systems, Nsight Compute, PyTorch Profiler
- Proficiency in Python, Rust, or Go
- Proficiency in C++ and CUDA
- Bonus: custom Triton kernels, diffusion/image model inference optimization, open-source inference framework contributions
Am I A Good Fit?
Get Personalized Job Insights.
Our AI-powered fit analysis compares your resume with a job listing so you know if your skills & experience align.
Success! Refresh the page to see how your skills align with this role.
The Company
What We Do
A cross-border crypto venture fund. Global from day one.
