How Zo Computer improved AI reliability 20x on Vercel
By Steven Van ·
Moving to Vercel's AI SDK and AI Gateway cut retry rates from 7.5% to 0.34% and trimmed P99 latency by 38%, Vercel says.
Zo Computer, a personal AI cloud platform, moved its backend to Vercel's AI SDK and AI Gateway to replace the custom adapter code it had written for each model provider it supports, including OpenAI, Anthropic, MiniMax, GLM and Fireworks. Vercel detailed the results in a case study.
- Retry rate dropped from 7.5% to 0.34%, a 20x improvement, with average attempts per chat reaching 1.00.
- Chat success rate rose to 99.93%, up from a 98% baseline.
- On MiniMax M2.5, Zo's most-used model, P99 latency fell 38%, from 131 seconds to 81 seconds.
- Adding support for a new model now takes about 30 seconds via a config string, down from an hour-long, multi-file code change.
By the end of the rollout, 91.88% of Zo's traffic ran through Vercel's routing, handling contexts averaging 42,500 input tokens versus 12,700 on the previous setup, at a lower error rate.
Vercel
The platform for frontend developers — deploy, preview, and scale web apps and AI agents with zero config.
View Vercel →