TixelJobs
P
Positron AIvia muse

ML Systems Engineer, Model Architectures (Mid/Senior/Principal)

REMOTE$160K - $350K/yrPosted 2d ago
ML EngineerStaff+Full-time#remote

About the Role

**About Positron AI**



Positron AI is building next\-generation AI inference accelerators designed from the ground up for low\-latency, high\-throughput large language model inference. Chips built for the training era keep running into two walls at inference time: memory bandwidth and chip\-to\-chip networking. We designed our architecture to break through both. Our first\-generation Atlas systems are already in production on Oracle Cloud Infrastructure. Asimov, our first custom ASIC, takes that architecture to silicon. In simulations, Asimov delivers several times more tokens per dollar and tokens per watt than leading GPUs, at a fraction of the power. Paired in our Titan system, Asimov is built to serve models beyond 16 trillion parameters. In September 2026, we raised $875 million at a $5 billion valuation. NEA led the round, joined by Atreides Management, Valor Equity Partners, and others, to bring Asimov and Titan to market. Our team includes engineers from Groq, Lambda, NVIDIA, Meta, F5, and many other industry leaders.

**Role Overview**
-----------------


As a ML Systems Engineer, you will be the engineer who gets new model architectures running on our hardware first, in simulation, before RTL or silicon exists. You will work with chip architects, the compiler and SDK teams, and customers to understand what each new architecture demands of the hardware and to make sure our simulators can execute it bit\-exactly.
The work is implementation\-heavy: new attention variants, mixture\-of\-experts, embeddings, and quantization formats, implemented in our simulator and verified against reference frameworks. You will set the standard turnaround for bringing up a new model and drive toward making that path fully automated.

**Key Responsibilities**
------------------------

**Simulator Implementation*** Implement the operators, numerics, and memory layouts that new model architectures need in our simulator, bit\-exact against a reference implementation.
* Extend simulator coverage across the full inference path, including prefill and decode, embeddings, attention variants (grouped\-query, gated, linear, and sparse), mixture\-of\-experts routing, and quantized formats.
* Maintain reference comparisons and regression coverage so model support stays correct as the simulator and hardware evolve.

**Architecture Evaluation*** Evaluate new architectures before hardware exists, identifying what they require of compute, memory, and data movement and what the chip would need to change.

**Customer Model Bring\-Up and Automation*** Bring customer models up in simulation ahead of emulator and hardware runs, and own the turnaround time for doing so.
* Partner with the agentic tooling engineer to automate new\-model bring\-up end to end, from model specification to simulator coverage.

**Required Qualifications**
---------------------------

* 4\+ years of experience as an ML systems or inference engineer.
* A strong understanding of transformer internals at the operator level, including attention variants, mixture\-of\-experts, embeddings, KV cache, and quantization (INT8 and FP8\).
* Demonstrated experience implementing or porting model inference by hand in PyTorch, llama.cpp, vLLM, TensorRT\-LLM, or a custom runtime.
* Hands\-on experience debugging numerical mismatches against a reference down to the operator and the bit.
* Ability to read a model architecture paper and turn it into a working, verified implementation.
* Strong Rust skills, or expert\-level C/C\+\+ skills and a willingness to work in Rust day to day, along with strong Python skills for model and reference work.
* Excellent communication skills, with the ability to explain what a new architecture needs from the hardware and where the current design falls short.

**Preferred Qualifications**
----------------------------

* Experience mapping models onto accelerator hardware or writing kernels for a non\-GPU architecture.
* Prior experience serving as the technical po
Share