LATEST04 AUG 2026 · 6 MIN
Why your AI demo breaks in week one
The three things that always go wrong the moment real users touch an AI prototype, and the small changes that stop them.
Read the post →generate.ts
// demo
await model(prompt)
// production
await withRetry(3, …)
if (!schema.ok) fallback()
ALSO PLANNED
What "production-ready" actually means for an AI feature
PLANNEDA checklist I run before I'll tell a client something is ready to put in front of real users.
MCP tool escalation patterns from production
PLANNEDThe three patterns I use when an MCP tool call fails, and when to escalate to a human vs. a fallback path.
Observability for LLM apps, without the vendor lock-in
PLANNEDWhat to trace, what to log, and how to know why a generation went wrong, using tools you already run.
Retries, timeouts, and the real cost of "just call the API"
PLANNEDA pattern for wrapping model calls so they fail quietly, recover on their own, and never surprise you on the invoice.
Calltura's voice-agent architecture
PLANNEDHow Calltura's telephony → STT → LLM → TTS loop holds up when calls don't wait.
New posts go out on X first. No newsletter, no popup.