Running the investment arm of a family on AI rails

The mandate

The investment arm of a family. A concentrated portfolio spanning public equities and private positions, run to a real objective with real money, not a model in a spreadsheet. Two jobs sat inside one mandate: produce results, and build the operating system that produces them reliably, so the returns come from a process and not from luck or a single good year.

The problem

A family portfolio at this scale runs into two failure modes. The first is drift: decisions get made in conversations, half-remembered later, and quietly contradicted. The second is noise: too much to monitor, too many holdings and moving parts for any one person to track by attention alone, so things get missed. A concentrated book makes both worse, because concentration means the mistakes are larger when they land.

The task was to run the money well and, at the same time, to remove the human failure points from how it gets run.

What I built and ran

Two things at once, which is the point of the case.

The first is the result. The portfolio has run ahead of the S&P 500 by roughly eight percent over two years, with the private positions tracking above the family's own target range. That result came from a concentrated, conviction-weighted book, not from spreading thin and matching the index.

The second is the machine underneath it. I built an AI operations layer that runs the day-to-day discipline of the portfolio so that judgment, mine, is spent on decisions rather than on data-gathering and monitoring.

How it works, the operating layer

Four parts, each doing a job a human does badly under load.

Research runs through AI pipelines that pull, reconcile, and re-derive the numbers behind every position, with a standing rule that any load-bearing figure gets recomputed from raw source rather than trusted from a summary. The pipeline catches the classic aggregator trap, a cumulative return quietly presented as an annualized one, before it reaches a decision.

Monitoring runs continuously against a defined scoreboard. The portfolio is measured against its benchmarks on a live basis, and the system is built to fail loud: if a data slot is empty, it refuses to produce a number rather than guessing one. A flag stays a flag until it is resolved, not until it is forgotten.

Decision hygiene is enforced in the records, not in memory. Every material decision is written down with its reasoning at the moment it is made, so the book carries its own audit trail. The thing that got decided and why is always recoverable, which is exactly what a portfolio run across years and across advisors needs.

Execution is gated. Nothing trades automatically. The AI does the reading, the scoring, and the monitoring; a human makes and authorizes every actual move, and that boundary is asserted in the system itself, not just as a policy. The infrastructure is there to make the human decision better and better-informed, never to make it unattended.

The outcome

The portfolio produces a result ahead of its public benchmark, and it produces it from a system that a second person could pick up and run, because the reasoning, the measurements, and the guardrails all live in the records rather than in one person's head. That is the difference between a good run and a repeatable one.

What this shows

I can run a real book to a real result, and I can build the AI infrastructure that makes that result a process rather than a story. The operating layer is the transferable part: research pipelines that check their own arithmetic, monitoring that fails loud, decision records that survive the people who made them, and execution that stays under human control by design. That is how I think AI belongs in serious operations, as the discipline underneath the judgment, not as a substitute for it.

One of a set of engagement and systems case studies. If a problem like this is on your desk, start with a conversation.