Deploying AI means versioning prompts, models, and graphs with the same discipline as any production API.
In my daily workflow
- I deploy behind feature flags or env toggles when changing models.
- I run smoke evals post-deploy before announcing changes.
- I watch error budgets and latency for assistant routes.
- I document rollback: previous model ID, prompt hash, graph version.
How it makes me work smarter
AI deploys fail quietly — quality drifts before errors spike. Pairing deploys with eval smoke tests and PostHog dashboards catches regressions early. Vercel previews let me test graph changes per branch.
My setup
- Vercel for Next.js API and server actions
- Environment-based model configuration
- Sentry release tracking
- Static generation for gallery; dynamic for assistant
On this portfolio
Gallery detail pages are statically generated; assistant and copilot stay dynamic with server-side keys. Admin model settings propagate through the provider router without redeploying hardcoded model strings.
