Senior Software Engineer — ML Inference & Numerics

Tensordyne; United States; Remote; USD 180,000 - 260,000; Full time

Job summary

Tensordyne is hiring a software engineer to optimize LLM inference on custom silicon. You will go from model code to hardware performance, building agentic harnesses, and ensuring repeatable experimentation. The role focuses on numerical optimization to balance precision, speed, and model quality.

You will own substantial parts of the SDK, work across models, compilers, runtime, and hardware, and drive architecture-aware software improvements for scalable inference on our accelerators.

Qualifications

• You understand LLM inference end-to-end, including prefill, decode, KV caches, and batching.
• You build robust, testable software with strong interfaces and abstractions.
• You are comfortable with agentic tooling and running multiple agents in parallel.

Responsibilities

• Build agentic workflows to inspect models, modify implementations, run experiments, and diagnose failures.
• Bring up new models across architectures for validated, high-performance execution on our accelerator.
• Optimize numerics: precision, accumulation, and quantization across kernels and memory.
• Make numerical quality measurable with tests comparing activations and outputs to references.
• Ship maintainable software with clear APIs and reproduce difficult bugs.
• Collaborate with hardware teams to evaluate numerical formats and inform design.

Skills

LLM inference

Python

Agentic tooling

Performance optimization

Distributed systems

Tools

ML compilers

Accelerator SDKs

Model compression

Job description

About Tensordyne

AI is transforming our world. It can perform cognitive functions that previously only humans could do, such as perceiving interactions across different modalities and environments - with the ability to quickly learn and then solve complex problems. Tensordyne is an AI system solution company that builds very high-performance, low-power generative AI inference systems. Our mission, through the creation of custom silicon, hardware and software, is to enable multimodal Generative AI inference acceleration at scale, with safe, sustainable, high-performance systems for our hyperscaler and neocloud data center customers. We are at the leading edge of advancing the latest research and product improvements for generative Al inference solutions that will make Al even more advantageous for compelling new generative AI applications.Tensordyne is a well funded, fast-paced startup company with headquarters in both Sunnyvale, CA, and Munich, Germany. We also have many talented team members working remotely across North America and Europe. We take care of our people and their families with comprehensive benefits, competitive compensation, flexible spending options, and recognition programs, because building category-defining technology starts with a healthy, supported team. Come join us as we shape the future of multimodal generative artificial intelligence!

The role

Help make the latest AI models run well on our silicon. You’ll work from model code to measured hardware performance: bring up new architectures, track down accuracy failures, and improve latency, throughput, and memory use.

The bigger challenge is making that work repeatable. You’ll build an agentic harness that can explore implementations, run experiments, check results, and guide the next change. Every investigation should leave behind better software, better tests, and a shorter path for the next model.

We’re looking for an excellent software engineer who understands LLM inference deeply and knows how to put coding agents to work. You’ll own substantial parts of our SDK and work closely with engineers across models, compilers, runtime, and hardware. Your particular focus will be numerical optimization: finding where precision matters, where computation can be cheaper, and how to preserve model quality.

What you’ll do

• Build the agentic workflows. Give agents the tools to inspect models, change implementations and numerical settings, run emulation and hardware experiments, and diagnose failures. Build the feedback and evaluation that make the results trustworthy.
• Bring up new models quickly; make them go brrrr. Work with engineers across the stack to take unfamiliar architectures from reference code to validated, performant execution on our accelerator. Turn what you learn into reusable model support and SDK improvements.
• Optimize numerics. Explore precision, accumulation, and quantization choices across the model. Understand how they interact with kernels, KV caches, memory, and distributed execution, then measure the gains in latency, throughput, and memory use against model quality on silicon.
• Make numerical quality measurable. Build conversion and emulation software that lets you test numerical choices quickly. Compare activations and outputs against reference models, trace accuracy regressions to their source, and validate changes with end-to-end model evaluations.
• Ship software that lasts. Own design, implementation, testing, and delivery. Build clear APIs that are easy to use for agents. Make difficult bugs reproducible and keep them from returning. Know when to build fast, and when to build thoroughly.
• Shape the next chip. Work with hardware architects to test how numerical formats and arithmetic choices affect real models. Turn those findings into better hardware and algorithm designs.

What we need from you

• You understand LLM inference end to end. You can read a new model’s implementation and reason about prefill, decode, attention, KV caches, MoE, batching, and parallelism. You understand how these translate into compute, memory, and communication on an accelerator.
• You build excellent software. You’ve owned substantial systems and delivered complex features from design through production. You’re strong in Python, thoughtful about interfaces, and rigorous about abstractions and testing.
• You are fluent with agentic tooling. You’d rather build the agent that runs the optimization than doing it by hand; and running numerous agents in parallel excites you. You know what is required for agents to work effectively in a large software organization, and you have the judgement when an agentic loop still needs a human.
• You take ownership and bring others along. You form clear hypotheses, test them, and drive problems to resolution across software boundaries. You communicate clearly through specs, design docs, and reviews that people and agents can act on. You explain complex trade-offs, give direct feedback, and make the people around you better.

If you tick a few of these boxes, we’d love to hear from you.

We’d be thrilled if you had

• Expertise in improving inference systems like debugging and optimizing the runtime of complex models and reasoning about accuracy-performance trade-offs.
• Experience with ML compilers, accelerator SDKs, or model compression libraries, particularly for non-NVIDIA accelerators such as TPU, Trainium/Inferentia, Tenstorrent, Groq, or Cerebras.
• Experience with model quantization, sparsity, codebooks, low-rank factorization or other compression techniques.
• Contributions to distributed inference or serving frameworks such as vLLM or SGLang.

Why join us

Your code will help determine how quickly we support new models, how efficiently they run, and what we build into our next chip. You’ll own meaningful software, work directly with the people designing the hardware, and see your decisions tested on real silicon.

If you want to build the tools, solve the hard failures, and help set the engineering direction of a new inference platform, we’d like to hear from you.

Tensordyne’s culture was built on the following values

• Put people first. We only succeed when our people succeed.
• Ethics and integrity always. Be open, honest, and respectful of everyone.
• Think Big. Be ambitious and have audacious goals.
• Aim for excellence. Quality and excellence count in everything we do.
• Own it and get it done. Results matter!
• Make each person better together than they would be as an individual.
• Embrace each other’s differences, and embrace that there will be differences.

Tensordyne is an equal opportunity employer. We believe diverse teams are better equipped to solve complex problems and build exceptional technology. All qualified applicants will receive consideration for employment without regard to age, color, gender identity or expression, marital status, national origin, disability, protected veteran status, race, religion, pregnancy, sexual orientation, or any other characteristic protected by applicable laws, regulations, and ordinances.