Three tactics help Higharc ship machine learning models faster
The handoff between machine learning researchers and software engineers is a notorious bottleneck. Researchers excel at exploring novel architectures, but rarely focus on writing production-grade APIs. Conversely, software engineers write highly robust systems but often lack the deep domain knowledge required to run complex computer vision or custom transformer models. At Higharc, Senior Research Engineer Vaidas Razgaitis solves this operational friction with a highly tactical systems-and-process approach.
1. Enforce Research Legibility with an RPT Document
To bridge the knowledge gap, Razgaitis introduces a standardized document: the Research Prototype Taxonomy (RPT). Mirroring traditional software engineering technical design documents, the RPT clarifies domain-specific concepts for engineers who might be unfamiliar with spatial reasoning, embedding models, or custom data structures.
By laying out the "why" behind the ML tool and establishing clear type contracts between the core product repository and the ML codebase, teams remove ambiguity. Crucially, the RPT identifies the persistence layer. Since ML researchers should not spend their time optimizing database storage, this section acts as the perfect handoff point where back-end engineers can step in to build production infrastructure.
2. Decouple Code with a Python Mono Repo
Higharc manages its ML code in a dedicated, Python-based mono repo composed of isolated, fully decoupled microservices. Rather than bundling all research projects together, the team aligns researchers to microservices on a near one-to-one ratio.
+-----------------+
| API Gateway |
+--------+--------+
|
+---------------------+---------------------+
| | |
+--------v--------+ +--------v--------+ +--------v--------+
| Microservice A | | Microservice B | | Microservice C |
| (Transformers) | | (Diffusions) | | (Vision Models) |
+-----------------+ +-----------------+ +-----------------+
Each microservice runs as a standalone application inside a unified Docker Bridge network, protected by an API gateway. By structuring these services around a simple layered architecture (API router, controllers, core business logic, and data layer), engineers easily wrap complex models using FastAPI and manage dependencies with Poetry or uv.
3. Slice Monolithic Prototypes Using Stacked PRs
Shipping a large, monolithic research prototype into production at once is a recipe for system instability and stalled reviews. Instead, teams should decompose these prototypes along distinct structural layers, using toolchains designed for progressive deployment.
Higharc relies on Graphite to manage stacked pull requests. This workflow lets domain experts asynchronously review specific slices of the codebase (such as type contracts, core business logic, or data models) without blocking overall developer momentum. This methodical decomposition makes shipping schedules highly predictable.
Research to Reality: Bringing Frontier ML Research to Production - Vaidas Razgaitis, Higharc
WatchAI Engineer // 14:57
We turn high signal in-person events for the top AI engineers, founders, leaders, and researchers in the world into the best free learning opportunities for millions around the world here on YouTube. Your subscribes, likes, comments, speaking, attendance, or sponsorships goes a long way toward making our biz model sustainable indefinitely. We strongly believe this industry deserves a better class of community and that we know how to do this well; we just need your support.