About the role

Own a large-scale foundation model end to end.
Design and implement custom CUDA kernels.
Architect and scale distributed training pipelines.
Build ML systems with strict production SLOs.
Work in a fast, ambiguous environment.

Skills the posting asks for

CUDA
C++
Python
PyTorch
distributed training

Summarised from the employer’s posting, which is reproduced in full below.

The employer’s full posting

About the company

A VC-backed AI/ML startup in West London building a novel foundation model for fully automated, unsupervised software delivery in embedded control systems. Early stage, high urgency, high transparency. They value directness over jargon and hands-on ownership over titles.

What you'll actually do

Own a large-scale foundation model end to end, from research through production. This is 0 to 1 work, not maintaining someone else's architecture.
Design and implement custom CUDA kernels where off-the-shelf libraries fall short.
Architect and scale distributed training and inference pipelines on cloud infrastructure.
Build and operate ML systems with strict production SLOs. Your models ship, not just train.
Create internal tooling and infrastructure that accelerates the whole team's output.
Work in a fast, ambiguous environment where you define what needs building next.

Core stack: CUDA, C++, Python, PyTorch, distributed training, GPU infrastructure, foundation models

Criteria I check in every CV (must-haves, only these matter)

You've shipped a large-scale foundation model from 0 to 1 at a high-growth AI/ML startup or a top-tier research lab. Papers alone don't count.
You've designed and implemented custom CUDA kernels to optimize model performance. This is non-negotiable.
You have direct hands-on experience scaling distributed training or inference pipelines on cloud infrastructure (AWS, GCP, or Azure).
You've owned an ML system with strict production SLOs or SLAs end to end, not just a research prototype.

What gets rejected immediately

Academic or research-only experience without commercial production delivery.
No proficiency in CUDA C/C++ or Python.
No distributed training or inference experience at scale.

EU AI Act compliance

I use TeamTailor's built-in Co-Pilot to extract signals from CVs during the review stage. The decision to move a candidate forward is always made by a human person (me or the client). No automated decisions are made about your application.