Professional Experience - Ali Mehrabian

Vector Institute, Toronto, Canada

Designing a debate-based multi-agent system and a prompt-engineered single-agent system for solving benchmarking questions in the automated capability evaluation (ACE) framework, using inter-agent critique, structured reasoning, and voting-based answer selection to improve performance on complex tasks in different domains.
Developed Echo-CoPilot, an agentic framework for healthcare-focused echocardiography video interpretation that orchestrates ReAct-style agents with knowledge-graph–guided measurement selection and self-contrast reconciliation, reducing hallucinated measurements and improving performance in clinical decision-making.
Developed MCPArena, an adaptive MCP-based framework that generates and benchmarks diverse tasks to evaluate tool-augmented AI agents across reasoning, retrieval, planning, and safety tasks.
Lead technical and organizational efforts for Vector's Cohort 10 (C10) and FastLane MLA Program, shaping ML project direction, roadmaps, and delivery with multiple industry partners.

Yale University, New Haven, CT, USA

Proposed SAMBA for stock-price prediction task, reduced RMSE by up to 10.5% and increased IC and RIC by up to 85.4% and 80.3%, respectively, across DJIA, NASDAQ, and NYSE stock data.
Advisor: Prof. Xiaohong Chen

The University of British Columbia (UBC), Vancouver, Canada

Supervisor: Prof. Vincent Wong
Proposed GNNs and Graph-Mamba models for cellular traffic prediction; improved RMSE and MAE by up to 15.50% and 7.44%; achieved up to 7× faster training and 681× fewer multiply-accumulate (MAC) operations.
Designed a heterogeneous graph-transformer network to solve resource-allocation optimization problems in RIS-aided MIMO-THz systems; improved system sum-rate by 8.8% and achieved faster training convergence.
Collaborator: Prof. Ilker Hacihaliloglu
Proposed a Fourier-based Kolmogorov-Arnold network for implicit neural representations; improved PSNR and SSIM on the image representation task by up to 8.91% and 5.62%, respectively, 11.12% for PSNR and 11.37% for SSIM on the image inpainting task, and IoU on the occupancy volume representation task by up to 0.96%.

SenseNet Inc., Vancouver, Canada

Developed and fine-tuned GNNs for wildfire anomaly detection from sensor data on AWS, with MLflow-based experiment tracking, CI/CD-enabled deployment, and a Dockerized FastAPI service for real-time model serving.

Sharif University of Technology (SUT), Tehran, Iran

Worked with Docker and SDN emulation tools to deploy 5G programmable network functions.
Supervisor: Prof. Babak Khalaj