About

I’m an AI generalist. I find practical problems worth solving, direct teams of AI agents to build the answer, and test what it can and can’t do before anyone relies on it. I also make films that explain AI products.

I came to AI from finance, operations and audit, and I still work in operations by day. As an audit intern at Deloitte in Abu Dhabi, I reconciled intercompany transactions and investigated payroll and journal-entry discrepancies for IFRS audits. At BluSky Restoration Contractors I coordinate project budgets, subcontracts, billing, margin reporting and weekly payroll. I received a company ROAR award, was featured in company communications within four months of joining, and was picked by the Regional Vice President to review time-and-materials workflows across the company.

That is also where I put AI to work. I showed AI how I sort a recurring report, and it wrote a script that now does that step in moments; I checked it against my own manual sort. AI also flags likely errors in weekly timecards, which I review myself.

Since May 2026 I’ve worked with AI agents every day, outside my day job.SourceSelf-reported
James’s own account. His local Claude Code and Codex logs show agent work on all 90 days from 30 July to 27 October 2026, counted by his local (Pacific) day; earlier logs are gone (Codex before 13 June, Claude Code before 21 July) and other tools are not logged, so May to July rests on his account. Since June that work has run through a system I built for it: one policy for every agent, separate worktrees for parallel builds, a shared memory, and review by a different model family at the gates that matter. My logs show agent work on each of the 90 days to 27 October 2026: 61.1 billion tokens across 5,560 sessions, most of it context the agents re-read.SourceMeasured at 8:25 pm Pacific on 27 Oct 2026 from James’s local Claude Code and Codex logs, by Pacific calendar day (30 Jul–27 Oct): 61.13 billion tokens processed, Claude Code 29.09 billion and Codex 32.04 billion. T3 Code’s usage dashboard, which reads the same logs by local day, showed 60.4 billion earlier that day. Agents re-read their working context on every turn, so 96.8% of these tokens were cache reads; the models wrote 225 million tokens of output. Earlier months and other tools are not counted. Since May my git identity is on 3,479 commits, most written by agents I directed.SourceUnique commits under James’s git identity since 1 May 2026 across 26 repository histories (copies of the same repository count once), deduplicated with git log on 27 Oct 2026. Agents commit under his identity, so most were written by agents at his direction.

With it I shipped Guardcell, an open-source Outlook add-in tested against pre-registered gates and 2,000 real emails, and published Receiver Reliance, a 1,805-episode study, then a follow-up that found no added benefit in the release decisions it tested. I directed the meeting studio Guardcell grew out of, and Catchwright, which I retired when its market didn’t hold up.

What I’m after is simple to say. Systems we rely on should be understandable, and provable. Built to help people, not only to save money. Open, so anyone can check. Real, rather than merely impressive.