Dreamineering

CV · The story

Fifty spreadsheets. One decision engine.

Before “forward deployed engineer” was a job title, I did the job at WorldCom. I worked inside the routing operation and turned how the work really ran into a system that ran it.

The work lived in fifty spreadsheets.

An international call can travel many routes. Each has a price, a quality and a limit. Some are bound by agreements between carriers.

The facts behind those choices sat in fifty spreadsheets: carrier rate sheets, number plans, contracts, capacity and call quality. They often disagreed.

When they disagreed, two things went wrong. Traffic took routes that looked cheap and failed. And competitors made money from the gaps. When new mobile number ranges reached one table before another, the same number could be priced two ways.

The problem wasn’t the maths. It was that nobody could trust the data the maths ran on.

We agreed what good looked like, then worked back.

The job was never “find the cheapest route”. It was to connect the call, at the agreed quality, inside the contracts, and still earn enough to keep the network worth using.

From that, the system followed:

  1. Agree what good looks like. Connect the call. Meet the quality floor. Honour the contracts. Earn enough to keep the network running.
  2. Build one source of truth. Rate sheets, number plans, contracts, capacity and quality, reconciled into one table everyone trusts.
  3. Rule out first. Drop any route that breaks a contract, lacks capacity, falls below the quality floor or sits on the blocklist.
  4. Rank what is left. Order the routes that pass by value, reliability and commitments.
  5. Act. Turn the ranking into routing plans the switches could run.
  6. Learn. Compare forecast with actual traffic, margin and call quality. Fix the next plan. Keep a record of why each route won.

Two results are on my CV. RFP responses went from days to hours, because four separate carrier hierarchies now lived in one tool. And the reference data stopped the arbitrage losses, because the model finally matched what the carrier agreements actually said.

Data analysis is finding where the truth disagrees.

Most businesses already have the data. It is spread across spreadsheets, inboxes and people’s heads, and the pieces don’t agree.

That is where I start. I work back from the outcome you want to where each piece of data begins. I mark where two sources disagree, where work waits, and where someone retypes the same thing. Then we agree the rules, put them in front of the decision, and measure what happens.

It works the same for a quote, an order, a booking or an invoice. A call route was just a decision made thousands of times a day.

Then it was people and spreadsheets. Now it’s agents and tools.

The loop hasn’t changed in twenty years. What runs it has. I now build the same loop with AI agents, on my own ventures first.

  1. 1Source of truth

    WorldCom, then: One governed routing table, built from fifty spreadsheets.

    Today: A shared knowledge base in plain text files. Decisions, rules and know-how live in one place, and every agent reads it before it acts.

  2. 2The steps

    WorldCom, then: A routing routine that people ran, with analysts preparing data and operators handling exceptions.

    Today: Atomic workflows: small steps in order, run by AI agents, with a person approving anything that matters.

  3. 3The checks

    WorldCom, then: Hard rules before ranking, and a quality floor no margin could buy.

    Today: Every workflow states what good looks like before it runs. Every run leaves a receipt of how it was checked.

  4. 4What repeats

    WorldCom, then: Rules encoded once in the system, not re-applied by hand in each spreadsheet.

    Today: Stackmates: shared tools for the work every venture repeats, so each new one starts further ahead.

Each run needs less of the expensive part.

AI is good at judgement, but judgement is costly and varies from run to run. So each time a step repeats, it moves down: from judgement, to a written skill, to a simple tool that gives the same answer every time.

The routing engine did this with people. The system applied the rules; people reviewed the exceptions.

The agreed picture of good is what makes that safe. Without it, you can’t tell whether the cheaper way is good enough.

What is proven, and what isn’t.

The routing work is twenty years old, and it ran in production. The agent version runs every day on my own ventures. It hasn’t yet run for an outside client. The first result will go up when it’s verified, with the owner’s say-so.

Bring one decision that keeps going wrong.

Tell me where the work should end and where the data comes from. We’ll find where the truth disagrees.

Email Nav

How working together goes · How we write a workflow · CV