Do you actually build this yourself, or just advise on it?
I build it. My background is finance — audit, fund services, private-equity fund finance — not software engineering. But the systems behind this site run for real: the capture pipeline, the daily brief that reads my own work back to me, the automations that file and tag everything. I designed and run them myself.
The method is simple — build fast with AI, then learn how it works by using and fixing it. I'm a beginner at the code and senior at the finance, and pretending otherwise on either side would be dishonest. Building it myself is the point: it keeps me grounded in what AI actually does, not what a demo promises.
What's your take on AI in finance?
AI rarely fails because of the model. It fails because of unclear ownership, inconsistent processes, and data that means different things to different people.
Most finance teams already know where the trouble is: the reconciliation everyone quietly re-checks by hand, the report that's manually adjusted every month, the workflow that would fall over the second you tried to automate it. Fixing those is just good operations — and it happens to be the groundwork that makes AI work. The hard part isn't getting access to clever tools. It's looking at your own operations honestly through that lens, and being willing to fix the plumbing first.
What is the Personal Operating System, and what do you build it with?
It's my own second brain and automation stack — capture what matters, process it into linked notes, and get a daily digest back — that I run for myself and write about here.
The stack is deliberately simple and portable. Everything lives as plain Markdown in an Obsidian vault, captured on the go through a Telegram bot, then processed and automated with Python and Claude, with local git as the backup. The shape is three moves — capture, process, digest — so nothing I save gets lost and the useful parts find their way back to me. I build in public because it keeps me honest: the journal records what actually shipped and what broke, not what was promised.
Will AI replace finance / accounting jobs?
It replaces tasks, not roles — and the tasks it takes are usually the ones you already resent: the reconciliation you re-key by hand, the report you rebuild every month. What it can't do is own the judgment call, carry the relationship, or have the instinct that says "this number looks wrong." The people at risk aren't the ones who use AI — they're the ones who keep doing by hand what it now does in seconds. So my honest answer is: get good at directing it, and you become more valuable, not less.
Everyone's piloting AI, but only a fraction of CFOs see real impact — why does it underdeliver?
Because the model was never the problem. AI fails on unclear ownership, inconsistent processes, and data that means different things to different people — the unglamorous foundations. Most pilots jump straight to the clever demo and never fix the plumbing underneath, so the AI just automates a mess faster. The few who get real impact did the boring work first: they agreed what the numbers mean, who owns them, and cleaned the inputs. That's not really an AI project — it's good operations that happens to make AI work.
Why is month-end close still a spreadsheet mess — can AI actually fix it?
The close is a mess because it's held together by people's memory and a web of spreadsheets only they fully understand — and that's a process problem before it's a technology one. AI can genuinely help: chase the reconciliations, flag the variances, assemble the file. But point it at an undocumented process and you just get a faster mess. Fix the close as a process first — write down the steps, agree the controls, standardise the inputs — and then automation has something solid to stand on. Most of the win is in that first, unglamorous pass.
We're drowning in disconnected systems (ERP, accounting, reporting). Where do you even start?
Not with the biggest problem — with the most repetitive one where the data is already clean enough to trust. Pick one high-volume, rules-based task you do every month — an AP match, a reconciliation, a recurring report — map it honestly end to end, and automate just that. One removed manual step you can see and verify beats a grand "transformation" that stalls. And keep it owned by finance, not handed to IT: the rules that matter — what counts as a match, what's an exception, what needs a human — are finance decisions. Get one win, learn from it, then take the next.
Is it safe to put AI anywhere near regulated / fund data?
It can be — but the burden is on you to make it so, and inside a regulated firm a lot of the popular tools simply aren't cleared for day-to-day use. The real question isn't "is AI safe," it's "where does this data go, can I stop it being used to train a model, and is this tool actually approved?" My own rule is local-first: keep anything sensitive on the machine, default to closed, and only send out what's genuinely safe to send. You can get a lot of value from AI without ever putting client data through someone else's model — and if you can't answer where the data goes, that's your answer.
Do I need to be technical / learn to code to use AI in finance?
No — and I'm proof you can go a lot further than "no-code" without being an engineer. You don't need to write code to use AI well; you need to know what it's good at, where it quietly gets things wrong, and how to check it. I started senior in finance and a beginner at the code, and I build by doing — fast with AI, then learning how it works by using and fixing it. The finance judgment is the hard part, and you already have it; the tools are learnable. Don't wait until you feel "technical enough" — start on a real problem you understand, and let the doing teach you.