Case studies

Examples of systems, projects, and ideas.

All work

Remodeling contractor

Estimate engine

  • Input: recorded site walkthrough
  • Classification over years of past jobs
  • Deterministic pricing in his own rules
  • Owner confirms every line before pricing

A contractor records the site walkthrough. The engine turns it into a complete estimate in his own line items and pricing rules, asks only what the data can't settle, and prices nothing until he confirms each line.

Mechanical contractor · Microsoft 365

Company AI Layer

  • Claude and Copilot side by side
  • SharePoint as the system of record
  • Live connections: accounting, time, vendors
  • Approval gate on anything that touches money

One structure under a company's existing tools. Each team gets its own space, instructions and skills; data from several vendors is read where it lives, never copied; a person approves anything that pays, sends or deletes.

Athletic department

Schedule system

  • Email and database kept in step
  • Integrated across several outside systems
  • Changes reach facilities, officials, transport, families
  • Downstream tasks reopen when inputs change

One master schedule. Move a game and the system knows what else has to change, drafts the notices, and holds the sends until a person approves them.

Installed, multi-user

Plugin workflow system

  • Packaged for Claude Cowork
  • Agents, skills and operations run by name
  • Connectors into the team's systems
  • Shared record store

A team installs one plugin and gets a set of jobs it can run by name. Each job is an agent with its own skills and rules, working from the team's own material and writing back to a shared record.

My own practice

Marketing research engine

  • Multi-agent pipeline on MCP servers
  • Market scan and competitor deep-dives
  • Positioning and a promotion plan
  • Every claim cited to its source

Point it at a market and it comes back with the landscape, the competitors, what to say and where to say it, with sources. Reruns weekly.

My own practice

Elysian command center

  • A director agent per engagement and function
  • Message bus between agents
  • Ticketed work with acceptance criteria
  • Independent review; approval before any external action

The system I run this practice on. Every agent has a defined job and defined permissions; nothing leaves the system without my approval. The same patterns go into every client build.

Any company with documents

RAG knowledge base

  • Retrieval over the company's own files
  • Chunk-level citations
  • Access scoped by role
  • Answers name the document and section

Ask a question in plain language and get the answer and where it came from: your safety manual, your job history, your contracts. Not the internet's version.

The connection layer

MCP servers

  • Custom Model Context Protocol servers
  • Expose job costing, CRM, scheduling as tools
  • Read-only first; write access is earned
  • Typed inputs and an audit log on every call

How Claude or Copilot reaches your actual systems safely: it can only do what the server allows, every call is logged, and nothing writes until it has proven itself.

The first conversation is free.

Start a conversation