Production AI is queues, retries, rate limits, on-call signals, and runbooks — not a notebook that happens to be deployed.
In my daily workflow
- I check Sentry and PostHog after deploys and on Monday mornings.
- I verify cron, webhooks, and assistant health endpoints.
- I keep a rollback checklist: model, prompt, graph, feature flag.
- I document incidents with eval cases to prevent repeats.
How it makes me work smarter
Production means predictable failure modes. Timeouts, graceful degradation ('assistant temporarily unavailable'), and idempotent writes turn demos into systems. The portfolio is live 24/7; treat assistant APIs accordingly.
My setup
- Vercel hosting with edge and Node runtimes as appropriate
- Rate limiting on public AI routes
- Sentry alerts; PostHog anomaly monitoring
- Keep-alive and CI smoke tests
On this portfolio
Assistant, copilot, and knowledge APIs run with observability hooks and cache rules suited to each route. The gallery is static for reliability; dynamic AI paths get the operational attention they need.
