Mate Security CEO Asaf Wiener shares how a spike in AI inference costs nearly killed the startup, prompting a full restructuring of how the company manages model selection. Rather than centralizing model decisions with a research team, every backend engineer now owns model evaluation and routing for their workloads. Engineers run continuous evals against frontier and open-source models, routing workloads to whichever model wins on quality and cost. Inference is broken into ~10 sub-lines on the CEO dashboard, measured per investigation. The approach also reshaped hiring — Mate filters for 'resilience for changing' and won't hire engineers uncomfortable with constant workflow churn. The result: overnight agent-generated integrations, sometimes self-approved, and a model layer that's no longer a fixed vendor cost.

11m read timeFrom thenewstack.io
Post cover image
Table of contents
Mate Security’s AI workflowWhen open source winsExecution mode, no refusers
463 Impressions