Artificial Intelligence Engineer

Acceler8 Talent; California (MO); Hybrid; USD 180,000 - 260,000; Full time

Job summary

Acceler8 Talent is seeking a Senior Pretraining Engineer to own large-scale video and multimodal pre-training. You will train models across distributed GPU infrastructure and push research into working systems for autonomous robotics applications.

The role focuses on video, vision, diffusion, flow matching and other generative-model approaches, requiring strong software engineering and training-systems fundamentals. Onsite in Bay Area.

Qualifications

• Large-scale video and multimodal pre-training
• Distributed GPU compute for large models
• Training large models on large datasets
• Debugging difficult training failures at scale
• Understanding interaction between models, data and compute
• Flow matching or diffusion models or adjacent generative-model approaches
• Strong software engineering and training-systems fundamentals
• Designing experiments where compute cost makes mistakes consequential

Responsibilities

• Own large-scale video and multimodal pre-training runs
• Train models across substantial distributed GPU infrastructure
• Develop and improve generative approaches including flow matching and diffusion
• Diagnose instability, convergence issues and large-scale training failures
• Improve training efficiency, reliability and experiment velocity
• Build safeguards based on lessons from failed or expensive training runs
• Work closely with robotics, simulation, perception and systems engineers
• Push research ideas into working systems rather than isolated experiments

Skills

Video pre-training

Multimodal models

Distributed training

GPU compute

Flow matching

Diffusion models

Training systems

Experimentation

Software engineering

Model debugging

Tools

PyTorch

CUDA

TensorFlow

Job description

Senior Pretraining Engineer – Video & Multimodal AI

Location:Bay Area, CA | Onsite

Focus:Large-scale pre-training, video, multimodal models, diffusion, flow matching

I'm working with an early-stage robotics company building a deeply integrated AI and robotics stack from first principles.

Their long-term goal is ambitious: create highly autonomous factories capable of manufacturing physical goods with dramatically less human manual labour. That means solving problems across robot learning, perception, simulation, rendering, GPU performance and large-scale multimodal model training.

They’re now hiring aSenior Pretraining Engineerto take ownership of large-scale video and multimodal pre-training.

This is not a role for someone who has only fine-tuned existing models or operated clean, established training pipelines.

You’ll be expected to understand what happens whenbig models, big datasets and big compute collide, including the failure modes that only become visible once training runs become genuinely expensive.

You’ll work across model architecture, training infrastructure, data and experimentation to build large-scale vision and multimodal systems that can ultimately contribute to robotic intelligence in complex physical environments.

• Own large-scale video and multimodal pre-training runs
• Train models across substantial distributed GPU infrastructure
• Develop and improve generative approaches includingflow matching and diffusion
• Diagnose instability, convergence issues and large-scale training failures
• Improve training efficiency, reliability and experiment velocity
• Build safeguards based on lessons from failed or expensive training runs
• Work closely with robotics, simulation, perception and systems engineers
• Push research ideas into working systems rather than isolated experiments

Strong candidates will have experience with:

• Large-scalevideo, vision or multimodal pre-training
• Distributed training across substantial GPU compute
• Training large models on large datasets
• Debugging difficult training failures at scale
• Understanding the interaction between models, data and compute
• Flow matching, diffusion models or adjacent generative-model approaches
• Strong software engineering and training-systems fundamentals
• Designing experiments where compute cost makes mistakes consequential

Just as importantly, you should be able to talk openly about training runs thatdidn’t work: what failed, how you diagnosed it, what it cost, and what you changed afterward.

The wider engineering environment spans:

• Robot learning and reinforcement learning
• Video and multimodal foundation models
• Perception for difficult real-world environments
• Custom physics simulation
• Rendering and light transport
• GPU kernel optimization
• Hardware-software co-design

That creates an unusually broad technical surface area. Your models won’t exist purely to improve benchmark scores. The longer-term objective is intelligence that can operate through physical systems and contribute to genuinely autonomous manufacturing.

The company is onsite in the Bay Area, with its R&D operation to be based in San Jose.

If you’ve personally taken large multimodal or video models through expensive pre‑training runs, including the painful ones, this is one of the more unusual opportunities to apply that experience to physical‑world AI.