Running tens of thousands of customer models
Thilo Huellmann, CTO of Levity, on using no-code AI for workflow automation
The real moat in no-code AI is not model quality, it is running huge numbers of small customer specific models cheaply enough for SMB budgets. Levity is selling to teams that want one classifier inside a live workflow, not an internal ML platform, so the hard part is keeping thousands of text and document models ready for unpredictable requests while still charging like software, not like a custom data science project.
-
Levity owns the full path from raw files to action. A user connects Gmail, Google Drive, or a CRM, pulls in past emails or PDFs, trains a classifier, then routes the prediction into the workflow. That end to end design makes model hosting and workflow uptime part of the product, not back office infrastructure.
-
The scaling problem comes from multitenancy at the long tail. A typical company might run a few dozen models, but Levity expects thousands of customers each deploying several niche models with bursty traffic. That is why it moved toward serverless inference and on demand training, to avoid paying for idle models all month.
-
This is also why Levity fits beside Zapier and looks similar to players like Nyckel. Zapier connects APIs but does not solve unstructured classification, while no-code ML platforms remove the need for an internal ML team. The winner is the one that hides data prep, deployment, and infra cost well enough that non technical teams can use ML as a routine workflow step.
As no-code AI opens from guided onboarding into trials and self serve, infrastructure efficiency becomes go to market leverage. The companies that can host millions of low frequency predictions, bake in native integrations, and reuse patterns across similar customer workflows will widen the gap, because cheaper deployment directly expands who can afford to automate.