Vaidas Razgaitis is a senior research engineer at Higharc who translates advances in machine learning into practical tools for homebuilding. His background as a structural engineer at Arup and an adjunct professor at Pratt Institute’s School of Architecture gives his work an architectural grounding that spans generative building design, spatial reasoning, and production software.
After working in structural engineering and architectural education, Razgaitis earned a master’s degree in computer science from the University of Chicago and moved into software development. His projects have included an architectural-design copilot, a social fitness application, and FAAS, a function-as-a-service platform. His geometry_engine performs three-dimensional geometric operations through a Python service, extending his interest in the computational representation of physical space.
At Higharc, he has worked on computer vision for hand-sketched floor plans, custom transformers, image generation, and reasoning systems. His contributions center on three connected problems:
- Generative Building Model: Razgaitis collaborated with Manuel Rodriguez Ladron de Guevara and Jinmo Rhee on a system that represents buildings through structured architectural elements. His approach to representing buildings like language preserves relationships among rooms, walls, doors, and windows, producing designs grounded in building geometry and semantics.
- Research-to-production handoff: He makes experimental systems understandable to engineers and product teams through technical documents covering architectural concepts, business goals, data representations, type contracts, persistence, and system boundaries. His production architecture combines isolated Python microservices, a Docker-networked gateway, layered FastAPI applications, automated testing, and dependency-aware code reviews.
- Adversarial coding agents: His model for self-healing codebases assigns one agent to identify weak code and propose repairs, another to evaluate those changes, and a moderator to expose gaps in their instructions. His example agent, Jacques-the-shrimp, uses failing tests, coverage analysis, corrective implementations, and automated pull requests, with human oversight and inspectable behavior remaining essential.
