Shipping ugly, how I got Maya live in 2.5 months

Most AI projects die in the gap between the demo and the deployment. The demo is easy now. Anyone can wire up something that looks like magic for five minutes. The deployment is where the weeks go, and where most projects quietly stall.

Maya did not stall. It went from nothing to a live customer support platform in about two and a half months. I want to write down how, because the honest version is not a story about being clever. It is a story about shipping ugly on purpose.

The clock and the constraint

I was the only AI person at ESC Partners when this started. No prior codebase to extend. No team to hand the hard parts to. Just a deadline and the whole surface area, a customer-facing assistant across voice, chat, SMS, email, and mobile, plus an internal console for the human agents.

A deadline that tight does you a favor. It takes gold-plating off the table before you can talk yourself into it. You cannot build the elegant version of everything in ten weeks. So you are forced to answer the only question that matters. What has to be right first?

What “ugly” actually means

Ugly does not mean broken. It means you ship the boring, high-volume path before the clever one.

Most support is repetitive. Bill questions. Outage status. Account changes. That is the bulk of what comes in, and it is not glamorous. So that is what I built first, and I made it solid before I touched anything rarer. One channel working end to end beats five channels half-wired.

The temptation is the opposite. Demo-brain wants you to polish the rare, impressive edge case, the one that makes people lean in during a meeting. That instinct is a trap. The wow-case is not what your customers actually send. Polishing it while the common path is still shaky is how you ship a thing that demos great and falls over on day one.

The hard part was the plumbing, not the model

Here is what surprised me least and cost me most. The model is the easy part in 2026. The hard part is wiring it into a real system of record.

The internal console plugs into Oracle CCS, so that when Maya hands a conversation to a human, the agent gets the full context loaded already instead of starting cold. That integration was the single hardest piece of the build. The documentation for that kind of system is not written for what I was doing with it. Nobody demos the integration. Everybody lives in it.

If you are scoping an AI product and your estimate is mostly model work, double the plumbing and you will be closer.

Shipping into real traffic

The moment Maya became real for me was not a benchmark. It was watching an agent use the console for the first time with a live customer on the line. They did not dig through three tabs to find the account. It was just there. That is the moment you cannot get from a sandbox.

Real traffic teaches you things no amount of pre-launch polish will. Customers phrase things you never anticipated. They ask three unrelated questions in one message. They show up upset on a billing cycle you forgot about. You learn more in a week of production than in a month of guessing.

The thing that let me ship ugly without shipping wrong was the eval suite. Speed is only safe if you have a way to catch yourself. I wrote about how I measure Maya honestly in The number I can defend. That gate is what kept “ship it rough and iterate” from turning into “ship it broken and hope.”

What I cut, and what I refused to cut

I cut plenty. Edge-case handling that could wait. UI flourish nobody would notice. The rare paths that were interesting to me and irrelevant to the customer.

I refused to cut three things. The honest handoff, where a human always inherits full context instead of a cold transfer. The eval gate, because shipping fast without measurement is just shipping blind. And the rule that nothing counts as resolved unless it actually was. Those three are not features you bolt on later. They are the spine. Cut them and the speed buys you nothing.

What the founder years taught me

I built two companies before this and sold both. The most useful thing I carried out of those years is not a framework. It is a reflex. Ship before perfect.

I made payroll when the bank account was thin. I watched good plans lose to shipped products, over and over. You do not learn what a product is by polishing the deck. You learn it the first time a real person uses it and does something you did not expect.

Maya in two and a half months is that reflex applied to AI. Ugly-shipped beats pretty-planned. The version in front of real customers is the only one that ever teaches you anything.