Production
AI
AI that survives contact with real users. A demo has to work once; production AI has to work on invoice day, at renewal, in Spanish, when the model is slow, and when it's confidently wrong.
What I mean by it
My bias here comes from operations. An AI feature in production is a dependency you can't page — so you design for its failure from day one: cached paths, degraded modes, and a human check on anything that faces a customer.
Exhibit A: Ancuria, live
Ancuria is a real-estate intelligence SaaS in production for the Mexican market — a bilingual PWA on Next.js 16 with AI where it earns its keep:
- ▸ Three-level AI assistant drafting property listings in Spanish and English, with human review before publish.
- ▸ RADAR calculators over 45+ daily market indicators — data pipelines that refresh themselves, not screenshots.
- ▸ Real billing: Conekta cards, SPEI bank transfer and cash payments, with CFDI invoicing — the part that proves a product is actually sold.
- ▸ Multi-tenant RBAC and a self-filling CRM.
ancuria.com · built under CreandoTuMatrix (Feb 2026–present)
The rest of the record
- ▸ Packt production-AI titles technically reviewed: "Microsoft Copilot in Azure" (2025), "Operational AI With Docker" (2026), "Microsoft Foundry in Action" (2026) — paid by a publisher to check whether other people's production-AI claims actually hold.
- ▸ MCP tooling connecting agents to real systems — from the 215+-tool environment integrated in Turbo-Flow to purpose-built tools — plus RAG pipelines in production contexts.
- ▸ Agentic operations: the same production standard applied to AI-built code — every agent change in my own stack passes an independent gate before merge (see agent governance).
Common questions
Where should a company put AI first?
Where failure is cheap and supervision is natural: drafting, summarizing, classifying — with a human approving output. Billing-critical decisions come last, after you have observability and fallbacks. That sequencing is exactly how Ancuria's assistant is built.
How do you keep AI features from degrading quietly?
Treat model output like a dependency with SLOs: log inputs/outputs, track cost and latency per feature, alert on drift, and keep a kill switch that falls back to the non-AI path. Production engineering, applied to a stranger dependency.