World's Fair 2026
Research to Reality: Bringing Frontier ML Research to Production - Vaidas Razgaitis, Higharc
Overview
This talk addresses the challenge of transitioning frontier machine learning research into production-ready software. It proposes a three-pronged approach focusing on research legibility, code structure, and decomposition strategies to improve the velocity and efficiency of teams bridging the gap between ML research and software engineering. The core idea is to treat this transition as a systems and process problem.
Who should watch
- AI Engineers
- Product Managers
- Software Engineers
- Infrastructure Engineers
- Builders looking to operationalize ML research
Key takeaways
- Establish clear documentation, such as a Research Prototype Taxonomy document, to make ML research concepts digestible for software engineers and product managers. This document should cover domain context, business goals, type safety, persistence layers, system architecture, and merging strategies.
- Structure codebases using a modular approach, such as a monorepo of cleanly isolated microservices, to facilitate the quick integration and iteration of new research prototypes.
- Implement a layered architecture within microservices (API, business logic, data layers) with well-documented specifications to aid navigation and acceleration by ML researchers.
- Decompose large, monolithic research prototypes into smaller, manageable units for easier review and integration into the production codebase. Tools like Graphite can aid in this process.
- The transition from research to production is a systems and process problem that can be improved by focusing on research legibility, code repository structure, and decomposition/PR review plans.
- Ambiguity in research legibility, difficulties integrating prototypes into the codebase, or challenges in estimating timelines for decomposition may indicate a need to revisit these processes.
Notable quotes
*The core idea is to treat this transition as a systems and process problem.*
*We need to start working with software engineers... who are very familiar with building robust and production-grade code, but are likely not familiar with the methodology and research in computer vision, in training your own LLMs.*
Unofficial community note. Prefer the recording for nuance.