Lead Research Scientist in London

Requirements

• Deep Research Expertise: PhD or equivalent deep industrial experience in Computer Science, Math, or Physics. You have a track record of publishing in top-tier conferences or shipping models that serve millions of users.
• Public Track Record: A portfolio of patents, impactful open-source contributions, or first-author publications in top-tier conferences (NeurIPS, ICML, CVPR, ICLR).
• Mastery of the Modern Stack: You are an expert in PyTorch or JAX. You can implement complex loss functions from scratch and debug distributed training issues on massive GPU clusters.
• Specialization in Structure & Generation: Deep experience in at least one of the following: Generative AI (LLMs/Diffusion), Graph Neural Networks (GNNs), or Geometric Deep Learning. You understand how to model relationships, not just tokens.
• Engineering Rigor: You write clean, modular, production-ready code. You understand the trade-offs between model accuracy and inference latency.
• Communication: You can explain the "Why" behind complex mathematical concepts to Product Managers, Designers, and Executives, turning abstract research into a compelling product vision.
• Option A: PhD in Computer Science, Machine Learning, Mathematics, Physics, or related field plus 4+ years of professional experience shipping ML at scale.
• Option B: Master’s degree or equivalent deep technical experience plus 7+ years of industry experience, with at least 2 years operating at a Senior or Staff level (driving technical strategy).
• (Desirable) Graph Learning Expertise: Specific experience with Graph Neural Networks (GNNs) or Geometric Deep Learning. You understand how to apply ML to non-Euclidean data (like the node-edge relationships on a Miro canvas).
• (Desirable) Generative Media: Hands-on experience building or fine-tuning Diffusion models for image/video generation or Multimodal LLMs (vision + text).
• (Desirable) Performance Optimization: Experience porting models to constrained environments (e.g., ONNX, WebGpu, CoreML) or optimizing inference for real-time interaction in the browser.
• (Desirable) Domain Knowledge: Previous work in Computational Creativity, HCI (Human-Computer Interaction), or building tools for thought (e.g., knowledge graphs, whiteboarding tools).

What the job involves

• Miro is looking for a Lead Research Scientist to serve as the technical "North Star" for our Machine Learning organization. You will operate as a high-level Individual Contributor (Staff/Principal level), driving the architectural decisions behind the "Intelligent Canvas." Your challenge is unique in the industry: You are not just processing text; you are building models that understand spatial relationships, visual diagrams, and unstructured collaboration.
• You will research, prototype, and ship novel architectures that combine Large Language Models (LLMs), Computer Vision, and Graph Neural Networks (GNNs) to make Miro the smartest collaboration platform on earth.
• Pioneer Novel Architectures: Move beyond off-the-shelf APIs. You will design and train custom architectures that fuse multimodal inputs (text, sketches, diagrams, screenshots, code, etc.) into a unified representation of user intent.
• Bridge Theory and Production: You will read the latest papers (NeurIPS, ICLR, CVPR) on Monday and have a working prototype by Friday. You bridge the gap between academic theory and scalable, low-latency production systems.
• Define the Technical Strategy: While Managers define what we build, you define how we build it. You will make high-stakes decisions on model selection (e.g., Diffusion vs. Autoregressive), build vs. buy, and fine-tuning strategies (LoRA, Q-LoRA).
• Mentorship & Technical Excellence: You will elevate the entire ML research engineering organization by conducting rigorous code reviews, hosting paper reading groups, and mentoring research engineers on mathematical fundamentals and experimental design.
• Solve the "Unsolved": You will tackle ambiguous problems with no StackOverflow answers—such as "How do we generate a valid UML diagram from a rough sticky-note brain dump?" or "How do we detect "agreement" in a spatial cluster of comments?"