About Tandemn
Tandemn builds an open-source planning and execution system for AI inference. It coordinates heterogeneous GPU clusters across clouds, regions, VPCs, and on-premises environments as a single pool of capacity. By planning workload configuration and placement together, Tandemn helps teams run existing workloads more efficiently and make room for new ones on the GPUs they already have.
Across evaluations, Tandemn delivers 2.4× effective GPU capacity and up to 92% lower cost per token. Our team brings together researchers, systems engineers, and mathematicians. You’ll work directly with the people building the system and own problems from definition through implementation and evaluation. Learn more about Tandemn.
The role
You’ll advance Koi, Tandemn’s fleet-wide planner, connecting workload requirements and hardware behavior to deployment decisions. This is hands-on research: you’ll define problems, implement methods, and evaluate them against real serving performance. The work spans algorithms and systems. We’re looking for exposure to both, with particular depth in mathematical modeling, optimization, or statistical learning.
Responsibilities
Build and improve the mathematical model of the fleet. The planner decides which workloads to admit, where to place them, and how to configure them. Because workloads share finite resources, these decisions must be evaluated together, and the number of possible plans makes brute-force search impractical. You’ll model inference behavior, develop optimization and search methods, and study how parallelism, batching, KV-cache use, and agentic workloads affect performance and cost. Design reproducible experiments and ablations in Python and PyTorch, use serving traces to test predictions, and work with systems engineers to put useful methods into the planning and execution loop.
Required qualifications
Minimum two years of ML research or applied research experience, including academic or industry work. Show us a problem you formulated, a method you implemented, and how you evaluated the result.
- Strong foundations in probability, statistics, optimization, and experimental design, with the ability to turn a mathematical idea into working code.
- Strong software engineering fundamentals: readable, tested Python, APIs, debugging, and practical experience with PyTorch and GPU experiments.
- Comfort with Linux, Git, and command-line workflows, and familiarity with processes, concurrency, containers, and networking.
- Familiarity with cloud and distributed systems, including VPCs, services, data models, serialization, and encoding.
- Exposure to inference engines such as vLLM, SGLang, or NVIDIA Dynamo, and techniques such as tensor and pipeline parallelism.
- Clear documentation, reproducible experiments, and a willingness to learn as models, serving systems, and workloads change.
Nice to have
- Experience with causal modeling, performance prediction, scheduling, or constrained optimization for shared resources.
- Kubernetes fundamentals, Helm charts, or orchestration tools such as Slurm or SkyPilot.
- Experience using Prometheus, Grafana, or serving telemetry to understand real inference behavior.
Traits we value
Treat AI-generated code as unverified. Use tools such as Claude and GPT to move faster, but review and test their output, understand the architecture and tradeoffs, and own the code you ship.
