How Zo Computer Made AI 20x More Reliable on Vercel

Zo Computer provides users with a personal AI cloud, running an always-on agent built on each user's servers, data, and files. With an 8-person team aiming to onboard 1 million users by 2026, reliably handling a massive volume of model calls became a core challenge.

Previously, the team had to maintain separate custom adapter code for each model provider, including OpenAI, Anthropic, MiniMax, GLM, and Fireworks. The team had to handle everything themselves—image processing, key management, exception handling, retry, provider routing, and fallback logic—meaning every new model required an hour-long multi-file change and deployment.

The turning point for Zo was the Vercel AI SDK and AI Gateway.

• The AI SDK replaced provider-specific adapter code, consolidating different implementations and edge-case handling for each model into a single unified interface.
• AI Gateway handled infrastructure-level complexity—retry, fallback routing, provider health monitoring, and uptime—at the routing layer.
• As a result, supporting a new model went from a 1-hour task to adding a 30-second config string, allowing Zo to bring MiniMax's M2.7 into production on the very day it launched.

The results were clear in the numbers. In an A/B comparison running Vercel and non-Vercel paths simultaneously, the retry rate dropped from 7.52% to 0.34%, a 20x improvement, and chat success rate rose to 99.93%. The average number of attempts became 1.00, while the non-Vercel path actually got worse over the same period.

Latency also improved significantly. For Zo's most-used model, MiniMax M2.5, P95 improved from 46 seconds to 34 seconds, and P99 improved from 131 seconds to 81 seconds. Average latency dropped by 25.7%, and by the end of the test, 91.88% of total traffic was routed through Vercel.

The Vercel path also handled larger contexts. Average input tokens were 42,500, 3.3x larger than the 12,700 of the non-Vercel path, yet the error rate was lower. Zo can now focus on its product instead of assembling infrastructure directly, and even with a small team, it can respond to model changes faster.