Research Scientist – Large Tabular Models (LTMs)

Research Scientist – Large Tabular Models (LTMs) Location: Mountain View, CA — Downtown HQ Employment Type: Full-time Work Model: On-site, 5 days/week Department: Research About Granica Granica helps enterprises run AI for less, from the data layer to agent execution . Granica is building across three related layers: Crunch reduces storage and processing costs across enterprise data infrastructure. Large Tabular Models unlock predictions and generated data from governed enterprise tables. Myelin reduces token and compute overhead for long-running AI agents. Everything runs inside the customer’s perimeter, so enterprises retain control of their data and the intelligence it creates. Granica has already processed hundreds of petabytes of enterprise data in production for customers including Snap, Pinterest, Lyft, and ShareChat . Our research organization is led by Stanford Professor Andrea Montanari . We operate at the intersection of data infrastructure, AI systems, and applied research , building the efficiency and intelligence layer for enterprise AI. Role Overview Most of today's generative AI is built for text, images, and video. Enterprise data isn't. The world's most valuable data lives in tables: customer records, transactions, financial systems, telemetry, operational data, and business workflows. Today's generative AI stack wasn't designed to learn efficiently from this kind of information. At Granica, we're building Large Tabular Models (LTMs) —foundation models that learn natively from structured and relational enterprise data. Our research is led by Prof. Andrea Montanari (Stanford) and focuses on one central question: How can we build generative AI that learns efficiently from tabular data? That requires solving problems well beyond model architecture, including intelligent data selection, dataset augmentation, representation learning, and information-preserving compression. If you're excited about inventing the algorithms that make Large Tabular Models possible, we'd love to talk. What You'll Work On Develop new machine learning algorithms for Large Tabular Models. Research methods for selecting, augmenting, and compressing training data without losing information. Build representation learning techniques for structured and relational datasets. Prototype and evaluate new approaches for generative modeling over enterprise data. Design rigorous experiments and benchmarks to measure progress. Collaborate closely with Prof. Andrea Montanari and Granica's research team to translate research into production systems. What We're Looking For PhD in Machine Learning, Computer Science, Statistics, Applied Mathematics, or a related field preferred; equivalent research and/or industry experience and demonstrated contributions to advancing state-of-the-art AI models will also be considered. Strong research record in machine learning. Experience developing new models or learning algorithms. Hands-on experience with PyTorch or JAX. Strong programming skills in Python. Ability to turn research ideas into working systems. Experience in structured learning, representation learning, generative modeling, probabilistic modeling, statistical learning, or scalable ML systems is particularly relevant. Bonus Research on tabular, relational, or graph data. Experience with diffusion or other generative modeling approaches. Publications at NeurIPS, ICML, ICLR, COLT, KDD, or related venues. Open-source or production ML systems experience. Compensation & Benefits Competitive salary, meaningful equity, and performance bonus for top performers 401(k) with company match, comprehensive health coverage, and unlimited PTO Daily catered meals in our Mountain View office Support for research, publication, and conference participation At Granica, you'll help build the next generation of enterprise AI —from exabyte-scale data infrastructure , Large Tabular Models (LTMs) , and stateful AI agents . Together, we're creating the infrastructure that enables enterprises to own their data , own the intelligence built on it , and scale both efficiently .