Descriptions:
Vaidas Razgaitis, senior research engineer at Higharc’s AI labs team, addresses a challenge familiar to any company running serious ML workloads: how do you move a research prototype — built by ML researchers comfortable with papers and experiments but not production-grade APIs — into a robust, maintainable production system built by software engineers who have never trained a model?
Higharc’s product spans computer vision for parsing hand-sketched floor plans, reasoning agents, custom transformer models, diffusion-based image generation, and more, making this handoff problem acute. Razgaitis organizes his solution around three levers. First, research legibility: every prototype must produce a “research prototype taxonomy document” — a TDD-style spec covering domain-specific data representations, business goals, type contracts, persistence layer design, system architecture, and a decomposition plan for production migration. Second, code structure: Higharc maintains a separate Python monorepo of cleanly isolated microservices, with roughly a one-to-one researcher-to-microservice ratio, each built as a layered FastAPI application (API routers → controllers → business logic → data layer) behind a central gateway. Third, production decomposition: the taxonomy document’s architecture section provides the roadmap for standing up each prototype, with agent-navigable documentation built in from the start.
The talk is practical and replicable, offering a transferable process for any team where ML researchers and platform engineers must collaborate without speaking each other’s language.
📺 Source: AI Engineer · Published June 28, 2026
🏷️ Format: Workflow Case Study
