Sina Shahandeh is co-founder and chief technology officer of RADiCAIT, where he is developing Insilico PET, a system designed to derive PET-like functional information from ordinary CT scans. His work on autonomous scientific research tackles a central limitation of AI agents: they can run experiments and improve code, but often struggle to generate the hypotheses that produce genuine scientific advances.
From materials science to medical imaging
Shahandeh studied materials science and engineering at Ferdowsi University of Mashhad and Sharif University of Technology before completing a doctorate at the University of British Columbia. His early computational research examined grain-boundary curvature and particle pinning, using mathematical models to investigate how microscopic structures evolve.
At smart-home company ecobee, he advanced from leading data and analytics to vice president of data science and AI. His work included machine learning for connected devices and collaboration on adaptive artificial agents for smart homes. He later served as founding chief technology officer at fertility-care company Ovom Care.
Shahandeh also founded Plan With Flow, a financial-planning application combining conversational AI with explicit models of income, expenses, assets, and liabilities. After the startup failed commercially, he open-sourced its code in 2025. He has separately proposed an ambitious agent-based simulation of the global economy, modeling interactions among households, businesses, institutions, and governments.
At RADiCAIT, Shahandeh applies generative imaging to pulmonary-nodule assessment and other clinical problems. Its proposed CT-to-PET approach aims to infer functional information usually associated with PET imaging without requiring a separate scan or additional radioactive tracer.
How he approaches autonomous research
Scientific hypothesis generation: Coding agents can exhaust familiar adjustments without questioning fundamental assumptions. Shahandeh structures research loops around proposing, criticizing, implementing, and evaluating new hypotheses—not merely repeating automated optimization.
Hierarchical problem decomposition: His research-agent framework maps data preparation, model architecture, training objectives, evaluation, and operational tooling into linked documents connected to implementation code. This structure helps agents identify consequential changes across an entire scientific problem.
Three-dimensional medical-image reasoning: A CT-to-PET model using stacked slices and two-dimensional convolutions can plateau when agents focus narrowly on hyperparameters. Explicit architectural analysis opens alternatives including three-dimensional convolutions and other representations that better capture anatomical structure.
Multimodal scientific evaluation: CT and PET scans may misalign because of breathing, patient movement, or differences between imaging sessions. Shahandeh combines numerical metrics with visual inspection of registration, lung masks, and cropping, while recognizing that current multimodal models still miss subtle features in specialized scientific imagery. Specialist models, adversarial critique, and additional test-time compute can strengthen the research loop, but dependable scientific observation remains a significant bottleneck.
