2025 · MICCAI
Saurav Bastola — AI researcher, Kathmandu, Nepal
I build models that learn how the world works.
I am a Research Assistant at NAAMII, working with Dr. Binod Bhattarai. My research is on generative world models and medical AI: unified, data-efficient internal models of the world, moving AI from pattern matching toward reasoning.
Currently I am developing diffusion models for spatial proteomics, generating synthetic biological tissue, and surgical world models that learn physics from unlabeled video.
Before research I built game AI at Bhoos Games. I studied Computer Engineering at Pulchowk Campus, and represented Nepal at the International Mathematical Olympiad in 2019.
- World models
- Generative AI
- Medical image synthesis
- Reinforcement learning
- Game AI
now
updated 2026-10
- Diffusion models for spatial proteomics (IMC) at NAAMII.
- Adapting GENIE-style generative interactive environments to medical imaging.
- Releasing RL Testbed, an open laboratory for learning reinforcement learning.
selected research
Publications →building
All projects →Reinforcement learning
RL Testbed ↗
An open laboratory for learning and doing RL: 29 algorithms from tabular Q-learning to world models and GRPO, implemented readably, with live training, seed comparisons and a learning path.
World models
Surgical Vision World Model ↗
A world model that learns surgical scene dynamics from unlabeled video. MICCAI 2025 workshop.
writing
All writing →- The Architecture of Discovery: Multi-Objective Diffusion for Spatial ProteomicsMy journey from blurry pixel-diffusion to a multi-objective Latent Model that learns biological plausibility through protein-to-protein translation.
- SurgWM: Learning Physics from Unlabeled Surgical VideoBuilding the first surgical vision world model that enables action-controllable generation without expensive action labels.
- From Scratch: Deriving Policy GradientsBridging the gap between basic MDPs and the math behind PPO/A3C. Why the 'Log-Derivative Trick' is the engine of modern RL.