cooperbuild.ai — the ontology layer for construction

the thesis

01

The harness, not the model.

Frontier AI is a commodity — Claude, GPT, Gemini, whatever ships next. The harness is what connects it to your business. We build the harness. You stay model-agnostic forever.

02

The ontology, not the document.

Construction firms don't have a data problem. They have an ontology problem. The objects exist — projects, subs, line items, RFIs — they're just trapped in PDFs and spreadsheets. We model them.

03

The deliverable templates that compound.

Before the estimate is even started, agents read canonical markdown that defines how the work should be built. Every closed job rewrites those templates. The wiki is the institutional knowledge — and it gets sharper with every project.

01 · the harness

AI alone is generic. Your data alone is dead weight. The harness is what makes them yours.

Cooper is model-agnostic by design.The frontier ships a new model every six months — we plug it in, your deployment gets smarter for free. You're not betting on Anthropic or OpenAI or Google. You're betting on your own data, modeled in a way every frontier model can reason over.

Frontier models

AClaude 4.5 Opusanthropic

OGPT-5openai

GGemini 2.5 Progoogle

LLlama 4 405Bmeta · self-hosted

?whatever ships nextmodel-agnostic by design

Cooper

The harness

Translates frontier reasoning into operations on your business graph.

Your business

Project1,284

Estimate3,612

LineItem412k

Subcontractor894

Specification28.4k

RFI11.2k

Schedule1,284

PurchaseOrder9,841

Model-agnostic · swap any frontier model without re-platforming

Self-hosted · SOC 2 · your data never trains a public model

02 · the ontology

Your business as a live, queryable graph.

Projects, line items, subs, specs, RFIs, schedules, POs — every object in your firm typed, related, and addressable. Not a data lake. Not a dashboard. An ontology. Click anything. Inspect its schema, its history, its relationships.

Object

Highland Retail TI

EST-0512 · active

Object

Estimate · v3

$2.32M · CSI-format

Object

Schedule

GANTT · 142 tasks

Document

Spec Set · 2024-09

412 sheets · indexed

Object

LineItem · 09 22 16.13

GWB · $84/sf

Object

Acme Drywall

8 jobs · 0.94 score

Object

PO-4421

$184k · open

Document

RFI-088 · slab

answered 2d ago

03 · the loop

A canonical wiki of how things should be built. Written by agents. Read by agents.

Before any estimate is started, agents read deliverable templates — markdown documents that encode the firm's institutional knowledge: pricing heuristics, sub-assignment policy, sequence rules, gotchas from past jobs. Every closed project writes back. The wiki is the source of truth, and it compounds.

deliverables/01-general-requirements/mobilization.md

Mobilization & Site Prep

canonicalrev 14 · 6 contributors

Mobilization & Site Prep

Canonical template — read before pricing any project under Div 01.

Scope

Establish site control, utilities, temporary facilities, and safety perimeter.

Pre-requisite to any sequenced work in Divs 02–33.

Inputs from ontology

• `Project.location` → drives permit set + AHJ

• `Project.duration` → temp facility burn rate

• `Project.client.type` → tenant vs. ground-up rules

Tasks (canonical sequence)

1. Pre-construction survey · 2-3 days

2. Permit acquisition · gated by AHJ (see local rules)

3. Site fencing + signage · per OSHA 1926.502

4. Temp utilities · power/water/data

5. Trailer + sanitation set · crew size driven

Pricing heuristics

• Burn = (duration_days * ${rate_by_region})

• Add 8-12% contingency for urban sites

• Hazmat review: trigger if `spec.contains("asbestos|lead")`

Known gotchas

• JFK FDA jobs: TSA badging adds 14 days, model into permit task.

• Highland Plaza portfolio: client supplies temp power, deduct $14k/mo.

the window

24 months.
The firms that model their business now own the next twenty years.

model capability

crossed

harness viability

live now

firms with an ontology

< 1%

window remaining

~24 mo

Frontier models crossed a threshold this year. Spec sets, RFIs, change orders — every unstructured artifact construction runs on — became queryable. The harness is finally possible. Two years ago it wasn't.

Most firms are paralyzed. They've watched the demos. They've hired a Director of AI. They've run six pilots that never made it past pilot. They don't have an AI problem. They have an ontology problem. The model has nothing to reason against.

The graph compounds. Every job makes it sharper. Every closed estimate, every RFI, every PO writes back to a substrate the laggards don't have and can't acquire. By the time they catch on, the gap is uncatchable.

04 · the agents

Queries, embodied as named agents.

Each agent is one query against the graph, given a job and a name. Tony estimates. Marcus schedules. Priya procures. Same substrate, different queries.

TM

Tony

Estimating · shipped Mar '26

live

the query behind tony

MATCH (p:Project)-[:GOVERNED_BY]->(s:Spec)
SELECT estimate(p, s, history)
WHERE p.id = "EST-0512"

05 · proof

Tony, in 12 seconds. One query against the graph.

Watch a real estimate compose itself from the ontology — spec → line items → bid → schedule recompute.

T

Tony

estimating agent · one query against the graph

objects produced · written back to ontology

06 · what cooper is not

Cooper is not — and that matters.

Positioning is sharper when the negatives are explicit.

01 · NOT

Cooper is not a vertical SaaS

We don't replace your project management, your accounting, or your ERP. We sit underneath them, modeling the objects they each see a slice of — so the graph is whole even when the tools are not.

02 · NOT

Cooper is not another AI copilot

Copilots wrap a model around a chat box and call it a day. Cooper is the substrate the model reasons against. The chat is incidental; the ontology is the product.

03 · NOT

Cooper is not a data warehouse

Warehouses store rows. Ontologies model relationships. We care about how a LineItem connects to a Spec, a Sub, a PO — not how to query a 12B-row fact table.

04 · NOT

Cooper is not a one-vendor model bet

Anthropic, OpenAI, Google, open-weight — the harness translates frontier reasoning into operations on your graph. The model is a swap; the graph is yours.

Construction runs on objects.
We model them. Any frontier model can query them.

The harness, not the model.

The ontology, not the document.

The deliverable templates that compound.

Model your business.
Before everyone else does.

The harness, not the model.

The ontology, not the document.

The deliverable templates that compound.

Model your business.Before everyone else does.

Model your business.
Before everyone else does.