AI Research Engineer, Computer Vision

Cantina

1d•$170,000 - $210,000

About The Position

We are looking for a talented AI Research Engineer to join our computer vision research team. In this role, you will work closely with our research team, implementing, training, and evaluating state-of-the-art image and video generation models. You will own the engineering execution that turns research ideas into working systems: building robust data pipelines, running and stabilizing large-scale training, implementing models from papers, optimizing for speed/efficiency, and running rigorous evaluations, all to accelerate our core image and video generation models. This is a high-impact implementation and execution role. This role is ideal for engineers who enjoy building reliable ML systems and scaling research ideas into production-quality training pipelines. The ideal candidate is someone who gets deep satisfaction from making complex systems work, translating research ideas into reliable, scalable code, debugging training instabilities, and delivering measurable improvements in training stability, model quality, and inference efficiency. This is an excellent opportunity to work closely with experienced researchers, gain deep hands-on exposure to cutting-edge model training techniques, latest research methods in diffusion/transformer-based generation, large-scale experimentation, and efficiency innovations, all while contributing directly to production-grade models.

Requirements

2–5 years of hands-on experience building and training ML systems, with strong ownership of results
Fluency in PyTorch: comfortable reading, writing, and debugging both training and inference code.
Experience training or fine-tuning generative models (diffusion models, transformers, VAEs, or similar) from scratch or near-scratch
Solid understanding of distributed training workflows and practical debugging of large training runs
Demonstrated ability to read and implement AI research papers in computer vision. Familiarity with cutting-edge computer vision models and research literature in the image and video domain.
Experience building data pipelines for large-scale image or video datasets
Strong debugging skills: comfortable diagnosing both engineering bugs and training failures
Strong engineering mindset: writing clean, reliable, debuggable code; profiling tools; handling numerical issues at scale.

Responsibilities

Build and maintain end-to-end data pipelines for large-scale image and video datasets: collection, filtering, augmentation, conditioning alignment, and efficient storage/sampling.
Implement model architectures (diffusion, autoregressive, flow-based, diffusion transformers, etc.) and maintain high-throughput PyTorch training loops for large-scale image and video diffusion models.
Run and manage large-scale training experiments on multi-GPU and multi-node setups (DDP, FSDP, DeepSpeed). Debug training instabilities, loss spikes, and convergence issues.
Apply quantization, pruning, and knowledge distillation techniques to compress models without sacrificing quality.
Collaborate with researchers and translate state-of-the-art research papers into working implementations in our internal codebase (e.g., new attention mechanisms, sampling schedules, or conditioning methods).
Build and maintain evaluation pipelines of image quality, video consistency, and perceptual metrics.
Set up and maintain human annotation and evaluation pipelines using services like AWS GroundTruth.
Profile and optimize training speed, GPU memory utilization, and iteration time. Implement inference optimizations to reduce latency and compute cost.
Work with acceleration toolchains such as torch.compile, Triton, TensorRT, or ONNX where appropriate