Required PhD Multimodality Researcher
Own the design, pretraining, and post-training of our multimodal foundation model for graphic design on multi-node GPU clusters.
About the role:
Graphic design is one of the last frontiers of multimodal AI that no single model has cracked end to end. It demands perception (what elements are on this canvas?), grounding (what is their type, hierarchy, and geometry?), reasoning about a latent design system (fonts, spacing, color, intent), and structured generation (emit an editable file, not a picture). Today we solve this by orchestrating specialized detection, inpainting, and code-generation models. You will lead the research to collapse that pipeline into a single multimodal foundation model for graphic design, and take it from a whiteboard to a training run to production.
In this role you will:
Design the model architecture, data strategy, and training recipe for a multimodal model that maps images to editable, structured design representations.
Run large-scale pretraining and post-training on multi-node GPU clusters, owning distributed training, sharding, checkpointing, throughput, and stability across nodes.
Build the post-training stack: supervised fine-tuning, preference optimization (RLHF and DPO style), and reward modeling grounded in real designer feedback.
Define rigorous evals for perception, structure recovery, and editability, and hold the model to them.
Ship. Your research reaches designers through the product, and their usage becomes your next training signal.
Requirements: You may be a good fit if you have:
A PhD in machine learning, computer vision, NLP, or a related field, or equivalent research experience with a track record of published or shipped work.
Deep, hands-on experience training large multimodal, vision-language, diffusion, or generative models from scratch.
Direct experience with multi-node, multi-GPU distributed training (for example FSDP, DeepSpeed, or Megatron) and the systems intuition to keep a large run healthy.
Strong software fundamentals in Python and PyTorch, and comfort living close to the metal when throughput demands it.
The pragmatism to care as much about the eval and the end user as about the loss curve.
Strong candidates may also have experience with
Object detection and UI or document understanding diffusion and inpainting structured and code generation with constrained decoding RLHF, DPO, or reward modeling data pipelines for web-scale image corpora inference optimization and quantization.
This position is open to all candidates.