Senior Research Engineer (up to £300k)

Dex; Greater London; On-site; GBP 90,000 - 150,000; Full time

Job summary

Dex is seeking a research engineer to join a core modeling stream, focusing on end-to-end development of proprietary generative models for photorealistic synthetic media. You will influence conditioning, distributed training stability, and inference performance across large GPU clusters, taking ideas from hypothesis through production.

The role emphasizes hands-on engineering over academic work, with responsibility from research design to production deployment in a fast-moving startup

Qualifications

• 6+ years of hands-on ML or research engineering.
• Proven track record training deep learning models at scale.
• Experience with multi-node distributed training across GPUs.
• Experience in training diffusion architectures in video or audio.

Responsibilities

• Own end-to-end research and engineering cycle for proprietary generative models, driving architectural hypotheses from small-scale validation to production deployment.
• Architect and execute distributed training recipes across multi-node, multi-GPU clusters under real compute constraints using FSDP, DeepSpeed, and sequence parallelism.
• Engineer conditioning mechanisms and training strategies that improve avatar control, temporal coherence, and motion fidelity without sacrificing photorealism.
• Isolate and resolve complex training instabilities, gradient anomalies, and throughput bottlenecks across large-scale model runs.
• Establish rigorous evaluation benchmarks and drive state-of-the-art inference optimizations to bridge the gap between offline model weights and low-latency product generation.

Skills

Python

PyTorch

Distributed training

Diffusion models

Research engineering

Tools

FSDP

DeepSpeed

Sequence parallelism

DDP

Job description

This role is with one of Dex's trusted partner companies. We work closely with their teams to truly understand their culture, goals, and what they're looking for, so we can match you with the right opportunity and give you context about the role before you commit to a process.

The role

While most generative video startups wrap third-party APIs or fine-tune open checkpoints, this company trains proprietary foundation models from scratch to power photorealistic synthetic media for tens of thousands of enterprise clients. Backed by hundreds of millions in funding from premier global tech investors, their lean in-house R&D team of roughly forty researchers and engineers operates at the frontier of video diffusion, real-time interactivity, and voice synthesis.

You will take on a high-leverage research engineering seat within a core modeling stream—spanning video pre-training, production post-training, interactive avatars, or audio modeling. This is not an academic post focused on publication metrics, nor is it routine ML infrastructure glue code. You will own the full arc from research hypothesis to production model, setting technical direction on conditioning mechanisms, distributed training stability, and inference performance across large GPU clusters.

The work

• Own the end-to-end research and engineering cycle for proprietary generative models, driving architectural hypotheses from small-scale validation to production deployment.
• Architect and execute distributed training recipes across multi-node, multi-GPU clusters under real compute constraints using FSDP, DeepSpeed, and sequence parallelism.
• Engineer conditioning mechanisms and training strategies that improve avatar control, temporal coherence, and motion fidelity without sacrificing photorealism.
• Isolate and resolve complex training instabilities, gradient anomalies, and throughput bottlenecks across large-scale model runs.
• Establish rigorous evaluation benchmarks and drive state-of-the-art inference optimizations to bridge the gap between offline model weights and low-latency product generation.

What You Bring

• 6+ years of hands-on experience in applied machine learning or research engineering, with a verified track record of training deep learning models at scale.
• Deep expertise in Python and PyTorch, alongside concrete experience operating multi-node distributed training clusters using FSDP, DeepSpeed, sequence parallelism, or DDP.
• Direct experience training diffusion architectures within temporal or sequential domains (video or audio generation; image-only experience is not sufficient).
• Proven ability to design rigorous, controlled ML experiments, interpret noisy training dynamics under real compute constraints, and discard dead ends quickly.

As part of the recruitment process at Dex, we process your personal data in accordance with our Privacy Notice for Job Applicants. This notice explains how and why your data is collected and used, and how you can contact us if you have any concerns.