Amir Tahmasbi

Proposed a two-stage framework that decomposes spatial reasoning into atomic building blocks and their composition. Applied supervised fine-tuning on elementary spatial transformations (rotation, translation, scaling), then trained lightweight LoRA adapters with GRPO to enable closed-loop, multi-step planning in puzzle-based environments.

Developing an interactive, story-based language-learning game using GPT that helps second-language learners practice slang and conversational English through spoken dialogue and reflection tasks. Designed as part of a study evaluating vocabulary retention among international students at Purdue University.

Collaborating with Lightspeed Studios on scalable traffic-flow simulation. Integrating Graph Neural Networks (GNN) with Multi-Agent PPO to derive environment-level coordination rules for complex multi-agent systems.

Path planning with Value Iteration and Random-based methods. Improved performance in complex maps by combining reinforcement learning with rapidly-exploring random trees.