cv

Founding Research Scientist - zenithon (May 2026 - current)

Developed real-time computer vision systems for surgical imaging. Worked on deep learning approaches to improve visual clarity in endoscopic video under strict latency and reliability constraints. Focused on ensuring outputs remain stable and faithful to the original scene rather than introducing artificial detail.
Led design of data annotation pipeline to produce model training data from raw surgical footage. Developed quality assurance tools and workflows to ensure high quality training data.
Introduced interpretable metrics for assessing model fidelity. Developed approaches to quantify how well models preserve important visual structures versus introducing artifacts. Emphasised decomposing model behavior into understandable components to better understand and mitigate risk.
Built interactive evaluation tools and human-in-the-loop workflows. Created lightweight interfaces to enable rapid visual inspection and comparison of model outputs. Incorporated user feedback and perceptual thresholds to better align quantitative metrics with human judgment.
Explored robustness to distribution shift in real-world settings. Investigated how variations in imaging conditions impact model performance and explored strategies to improve generalisation across environments. Focused on approaches that do not rely heavily on labeled data in new domains.

Developed generative models for 2D artists. Worked on models to map arbitrary artistic inputs (sketches, colour fills, etc.) to outputs conditioned on language-based prompts. Shipped multiple models to production.
Built internal evaluation framework to test model robustness across modalities
Explored techniques for fine-grained guidance of model outputs
Contributed fast sampler code to production, balancing image fidelity and sample spee

Implemented and debugged neural operator models for diffusion and super resolution-based weather forecasting, and open sourced code

Led and published three projects: 3D-inspired visual question answering; few-shot learning with synthetic data; and generative models for inverse design.

Explored image domain translation techniques for translating between sick and healthy modalities