case-studies / ancuria
Case Study: Building and Operating an AI SaaS in Production
The question every AI consultant should be asked: "Have you actually run this stuff in production?"
This is the answer. Ancuria — a real-estate intelligence SaaS for the Mexican market — designed, built, shipped, and operated, with the failures included.
What was built
- Payments that respect Mexico: card, SPEI, and cash through Conekta, with CFDI invoicing handled correctly — the part that quietly kills most "we'll expand to Mexico" plans.
- A three-level AI assistant: drafts bilingual (ES/EN) listings from partial input, and answers client chats — scoped, monitored, and bounded, per agent governance.
- RADAR calculators: financial tools over 45+ daily market indicators, refreshed on schedule.
- A CRM that fills itself: agent activity captured as a side effect of selling, not as data entry.
- Platform shape: bilingual PWA, installable, dark mode, web push; free tier with unlimited listings, paid plans $249–$699 MXN/month.
How it is operated
The interesting part of a production SaaS is not the feature list; it is what happens at 3 a.m. Ancuria runs on managed infrastructure with point-in-time-recovery database backups, scheduled data and maintenance jobs that report their own outcomes, and an operational rule borrowed from the rig: every automated process writes to a log a human can audit.
Incidents get runbooks. The AI features live under cost ceilings and scoped permissions. Nothing customer-facing ships without passing the same review gate as everything else in the ecosystem.
What broke, and what it taught
Production taught the lessons this site sells, the honest way:
Local success is not production success. Everything that worked in development got re-verified in production conditions — and some of it did not survive the trip.
The AI assistant needed boundaries, not more intelligence. Scope creep in agent behavior is a governance problem; fixed permissions and clear escalation beat prompt engineering for reliability.
Money paths are the hard part. Payments and invoicing consumed attention disproportionate to their code size — exactly why an adversarial architecture review before building is cheaper than after.
Observability paid for itself first. The logs that answered "what did the system do" were the cheapest component and the most valuable.
Why it strengthens the consulting
Because the recommendations are load-bearing. When the audit says "bound the permissions before the agent touches customers" or "make the payment flow reviewable end to end", that is not theory — it is scar tissue from a system where those exact calls were made, some right the first time and some not.
The system page with the full feature inventory: Ancuria.
Talk it through with someone who runs this stack on his own systems every day.