DWG. LL—01 Designer & Builder P-03 / IN BUILD Rev. 2026
← Index

Part P-03 — In build

Cosmo

An AI operating system for venues that run their business on systems that don't talk. One brain over the whole operation, agents that do the work.

RoleFounder
StatusIn build
InstanceOne live venue
StackPython · LangGraph · dlt · DuckDB
Overview

The venue finally talks to itself. Then Cosmo does the work: the quote, the booking, the schedule.

The problem

A venue that does several things at once runs a system for each of them. The point of sale, the events pipeline, the inventory, the books. None of them talk. Answering one question means five tabs, and acting on the answer means retyping it into a sixth. The gap between seeing something and doing something about it is where the money leaks.

FIG. 01 / Systems that don't talk
POS Events Inventory Books
Four systems · one question · five tabs
The shape of it

Cosmo is two layers. The brain syncs every system into one store you can ask questions of, with real definitions behind every metric. The agents sit on top. They read the brain, then do the work inside the venue's own systems: build the quote, confirm the booking, set the schedule.

Where a system has an API, the agent calls it. Where it doesn't, the agent works the app the way a person would. Either way, the same rules apply.

FIG. 02 / System schematic
The systems
POS Events Inventory Books
The brain
One brain Every system, queryable
The agents
Leads Quotes Bookings Schedules
Agents act back into the systems · every irreversible step gated

The rules it's built under

Rev. notes
01

The AI never computes a number

Every price, count, and total comes from deterministic code over real data. The model decides what to do next. It never does the math.

02

Every irreversible step is gated

Before anything that can't be undone, the agent acts, proposes for approval, or escalates to a person. One material error and it pauses itself.

03

Grounded, then verified

Every fact an agent uses comes from the brain, and the agent reads the result back from the system before it commits. Trust comes from the wrapping, not the model.

04

Judged on the hardest cases

The eval set is the cross-system questions that are easy to get wrong, not the easy ones that pad a pass rate. A system that's right on easy questions and wrong on hard ones has failed.

The hard part

Getting an agent to do the work isn't the hard part. Getting one you'd trust to send a quote or confirm a booking, without watching it, is. So most of the build is the part that isn't the model: the gates, the grounding, the read-back, the eval set. I treat the model as the least-trusted component in the system, and wrap it so it can't do damage on its own.

Where it stands

The brain is built and running on a real venue's data. The action layer is in build: the safety substrate first, then each capability behind the same gate. This sheet is the drawing, not the product. It updates as the build does.

Talk about the build → ← Back to index