Surya R
Building theoretical foundations and scalable architectures for next-generation intelligence— where pure mathematics meets applied machine learning.
Researching neuro-symbolic systems, efficient inference, and the boundaries of generative models. Based in the United Kingdom.
Scientist working at the intersection of theory and applied AI.
A short note on background, current research direction, and the kind of problems that get me out of bed.
I'm Surya — a scientist and AI researcher focused on the mathematical structures that make modern intelligence systems robust, efficient, and interpretable.
My work sits between deep learning research and production-grade systems: training foundation models, diagnosing why they fail, and engineering the infrastructure that lets them scale. I care about rigor, reproducibility, and ideas that survive contact with real data.
Outside the lab I read papers, write notes, and ship small tools that test hypotheses faster than they can be argued about.
Tools, methods, and territories I work in daily.
A working matrix of frameworks, mathematical foundations, and systems concerns.
Upcoming AI projects on the bench.
A working set of research threads I'm developing — some early, some closer to publication.
Neuro-Symbolic LLM Grounding
A framework integrating symbolic logic constraints into large language model decoding to eliminate factual hallucinations and enforce structured reasoning.
Federated Vision-Language RL
Edge-optimized reinforcement learning environments for autonomous multimodal agents operating across decentralized devices.
Generative Time-Series Imputation
Diffusion models applied to multivariate financial and medical time-series data with irregular missing intervals.
Scalable MLOps for Sparse Networks
Orchestration pipeline for training and deploying highly sparse neural networks without hardware degradation.
Selected publications & preprints.
A short editorial list. Work spans symbolic reasoning, vision, and probabilistic methods.
- 2025Attention is Not Enough: Symbolic Grounding in Autoregressive ModelsNeurIPS (Under Review)Conference
- 2024Efficient Edge-CV: Distilling Vision Transformers for Mobile NPUCVPR Workshop on Edge AIWorkshop
- 2024Probabilistic Forecasting under Concept DriftPreprint · arXivPreprint
- 2023Sparse Reward Shaping for Multi-Agent CoordinationICML Workshop on RLWorkshop