About
ML Research Engineer at CoSTAR National Lab working on video and world models for virtual production. Previously research intern at UIUC Blender Lab (energy-based models). MSc AI (Distinction) from the University of Essex, supervised by Prof. Luca Citi.
Research Interests
A representation of a system can be learned and refined from partial views and measurements. A signal, a geometric system evolving over time, or an image could be the different faces of the same underlying system or its evolved state. Most of what we ask of such a system, whether we are recovering something hidden, reconstructing what we cannot observe directly, or anticipating how it will behave, is the same question asked through a different measurement.
I'm interested in developing differentiable methods for learning such representations.When a general method can absorb the structure of a problem from scale, it usually overtakes the carefully engineered alternative .
Taken to its limit this points to one model rather than many. A representation of biological systems wide enough to reach across scales, from molecular structure through anatomy and imaging to how a patient changes over time, learned together so that what is known at one scale sharpens inference at another.
▸ Key areas Click to expand
- One representation, many questions: Treating tasks as conditionals of a single learned model of the system, where generalisation comes from a representation made rich by varied measurements and data modalities rather than from composition built in by hand.
- Continuous-time generative models: Diffusion and score-based models, flow matching and stochastic interpolants, Schrödinger bridges, optimal transport, and energy-based models, used as learned priors and samplers. Built on ODEs and SDEs, Fokker-Planck and Itô calculus, and probability-flow ideas.
- Amortised and simulation-based inference: Neural posterior estimation, amortised variational inference, and likelihood-free methods that turn expensive per-case fitting into a single forward pass.
- Learned priors for ill-posed problems: Expressive generative models that make sense of measurements which do not determine a unique answer, with calibrated uncertainty and correct coverage.
- Geometry and operators: Riemannian geometry and shape spaces for representation, and links to statistical mechanics.
- Generalisation and robustness: Out-of-distribution behaviour, transfer across systems and regimes, and reliability when data are scarce or noisy.
- Applications: Scientific inference, with digital twins and cardiac modelling as the present testbed.
Current Focus. Bringing diffusion, Schrödinger bridges, flow-based, and energy-based models under one view through statistical mechanics and differential geometry.
Writing
New Video World Models — From slow bidirectional diffusion to real-time autoregressive world simulators. Causality, self-forcing, attention sinks, and the path to interactive video generation.
News
- New Joined CoSTAR National Lab as ML Research Engineer - Feb. 2026
- Research Collaboration with UIUC Blender Lab - Jul. 2025
- MSc defended at Essex (NIR thesis) - Oct. 2024
Awards
- Full scholarship for MSc AI, University of Essex
Education
University of Essex
MSc Artificial Intelligence (Distinction)
2023 - 2024
Dissertation: NIR — Neural Integration of Iterative Reasoning in LLMs for Code Generation
🥇 1st place, Inter-Departmental Neural Network Challenge (100+ participants)
University of Kurdistan
BEng Computer Engineering
2014 - 2018
Thesis: EfficientCoF — Subspace clustering for collaborative filtering
Experience
CoSTAR National Lab
ML Research Engineer
Feb 2026 - Present
Generative video and world models for virtual production. [Profile]
UIUC Blender Lab
Research Intern
Jul 2025 - Jan 2026
Energy-based transformers for image/video generation. Investigating mode collapse, inference-time compute scaling, and fast sampling strategies.
University of Essex
Postgraduate Researcher
Apr 2024 - Oct 2024
Developed NIR framework for integrating iterative reasoning into LLM hidden states without fine-tuning.
Open Source
Developer & Maintainer
2019 - Present
TorchEBM, cuRBLAS, TransformerX, Emgraph, Bigraph, TASE. See Projects section.
Projects
TorchEBM 🍓
Updated Read more
PyTorch library for energy-based models, diffusion, and flow matching. Implements samplers, score/flow/contrastive objectives, SDE/ODE integrators, interpolants, and mixed-precision training.
PyTorch · CUDA
WorldKernels 
Soon High-throughput world model inference engine. Stateful session management for BiD and AR world models (DreamDojo, Cosmos, DreamZero, etc.) with persistent KV-cache, CUDA graph capture, continuous batching, speculative decoding, and torch.compile fusion. Sub-millisecond scheduling via async token queues; REST/WebSocket streaming with backpressure.
PyTorch · CUDA · Python
cuRBLAS 🍒
GPU-accelerated randomized linear algebra. CUDA kernels for Hutchinson trace estimation, randomized SVD, and probabilistic matrix operations—useful for sliced score matching and large-scale ML.
C++ · CUDA · Python
TransformerX
Modular transformer research library. Composable attention mechanisms, positional encodings, and architecture variants for rapid prototyping.
TensorFlow
Emgraph
Knowledge graph embedding library. Train and evaluate TransE, DistMult, ComplEx, and other relational models on link prediction tasks.
TensorFlow 2
Bigraph
Bipartite graph link prediction. Implements Jaccard, Adamic-Adar, Common Neighbors, Preferential Attachment, and Katz similarity for two-mode networks.
NetworkX · NumPy
Large-scale audio search engine. Scalable indexing and retrieval with Elasticsearch, ArangoDB, and Redis for real-time multimodal search.
Elasticsearch · Python
