Daniel Svonava is cofounder and CEO of Superlinked. A former machine-learning tech lead for YouTube Ads, he works on infrastructure for running small, task-specific models in AI applications.[1]
Start with Large clusters for small models →
Links
Superlinked LinkedInCofounder and CEO · Superlinked
Career
Svonava previously worked as a machine-learning tech lead for YouTube Ads at Google. Superlinked began in 2021 around search and recommendation systems. Its work with embeddings, rerankers and document models exposed the infrastructure needed to operate those components, leading to the Superlinked Inference Engine. The company now offers both a hosted API and deployments that teams can operate in their own environments.[1]
Contributions
SIE supplies serving infrastructure for embeddings, reranking, extraction and other model tasks. Its open-source stack supports local use and clustered deployments, with shared APIs and SDKs. The design lets teams use the hosted service or bring the same request format into their own cloud. The choice is about operating the infrastructure and controlling where requests run; model quality still has to be evaluated for the intended task.[3][4]
Ideas
In his conference talk, Svonava argues that an agent can use a fleet of smaller models rather than route every operation through one general model. The workflow begins by separating the application into tasks, evaluating candidates against those tasks and adapting a model when needed. This changes the infrastructure problem: a team has to manage different model families and workloads while preserving a consistent interface for the application.[2]
Tools, writing & talks
Writing, research & talks
Start here
Large clusters for small modelsai.engineerExplains evaluation-led model selection and the infrastructure needed for a fleet of specialized models.
Start with the example of splitting an agent workflow into distinct model tasks, then follow the evaluation process. Basic familiarity with model APIs is enough; deployment details are easier to follow after that decision is clear.
Speaker[2]Superlinked's Daniel Svonava on serving many small, specialized models instead of one big one: a shared-queue gateway that lets workers self-batch doubled cluster throughput versus router-based scheduling.
Catalog workOriginal source- Superlinked launchsuperlinked.com
The company’s account of its approach to small-model inference.
Company release[4]
Tools
- Superlinked Inference Engine
github.comOpen-source infrastructure for serving several model types through a common interface.