Signals before Headlines

The Kitchen Table Story — Spring 2026

At 57, I wake up early — earlier and earlier, actually. Usually around 5am, with a brain that's already running and a body that hasn't quite agreed yet.

I'd been watching the situation in Iran escalate for months — not as an analyst, not as a trader, but as someone with friends in the region. People I care about. And I kept finding myself refreshing feeds at odd hours, trying to piece together what was actually happening versus what the news was saying twelve hours later.

The gap between those two things bothered me. A lot.

One morning I asked Grok if a tool to extract signals from the news noise on SOCINT sources already existed. It said it didn't — but then it said: try Make.com. I had never heard of it.

My only coding experience was at 16 — Basic on a Commodore 64 I got for Christmas. That's it. Forty years of gap, and then Claude, Grok and ChatGPT. By the standards of the AI world, that makes me a dinosaur. Most people building with these tools are half my age, move faster, and probably never had to look up what a datastore was. I did, more than once.

I spent a whole spring on it, nine hours a day, talking to Claude like he was my late-night business partner. Sometimes we argued. Sometimes he surprised me. Other times he made me feel like an idiot because the prompt I'd written at 3am was generating complete nonsense.

There were nights when everything seemed to work: tweets coming in, the AI classifying topics, the system understanding that an event wasn't just being mentioned for the first time — it was changing state. Hormuz going from "closed for exercises" to "open." Tensions heating up and then cooling down. When I finally got that architecture right, when I added the state-change detection and the skip logic, it felt like removing a stone from my shoe that had been bothering me for weeks.

And then there were the mornings after.

Duplicate records everywhere. Signals being written identically every three hours. GPT-4o-mini cheerfully declaring a direction was "UP" because two tweets were optimistic in a sea of twenty catastrophists. Back to square one, swearing in Italian and English.

The hardest moment was realizing that "first time we see this topic" wasn't enough. The world isn't made of events that appear and disappear. It's made of things that change. I spent days rebuilding the architecture just to capture those direction shifts.

When it finally worked, I felt less like a builder and more like someone who had finally asked the right question.

By the end of that spring, the system actually existed. It monitored SOCINT sources on X, calculated how far ahead it saw state changes compared to mainstream coverage, checked whether the market moved in the expected direction within 48 hours, and built a track record. It's not magic. It's a lot of broken things, fixed slowly, one late night at a time.

The AI does the heavy lifting. But the most human part — deciding what's worth following, when to reset a lead time, when to trust a signal and when to ignore it — that part I still have to do myself.

And honestly, I wouldn't want it any other way.

I don't know yet what this becomes. Maybe something, maybe just my personal tool for trying to understand a world that feels increasingly hard to read. I'm not sure it matters right now.

What I know is that I built it — a 57-year-old with a Commodore 64 as his only coding credential, in an industry that sometimes feels like it was designed for people who've never had to worry about friends in a war zone. And every time I see a new signal with its lead time and its 48-hour verification, I feel like I'm accumulating something real: not a prediction, but a trace of how early certain voices on the web can smell the changes that end up moving markets.

This is the kitchen table version: a guy who spent a month talking to an AI like it was a friend, who cursed at datastores, who rebuilt his logic three times because he'd been asking the wrong question, and who ended up with something that actually works.

Or at least — it's starting to seem that way.

The Rewrite — Summer 2026

Make.com worked. That was the problem.

It worked well enough that I stopped thinking of it as an experiment and started thinking about a second user. Not someday — now. Someone paying me for it.

That's when the whole thing fell apart in my head. On Make, a second user meant duplicating every scenario by hand. A third meant doing it again. And if that same user wanted to follow two subjects instead of one, it was another full copy — the machine had no idea what a "subject" even was. It only knew the one it had been built around. Every question cost the full price of the first, over and over. I did the math on what a single customer would actually cost me to run and the number was worse than anything I could reasonably charge. A business where every new client makes you poorer isn't a business. It's a hobby that sends invoices.

So I threw away a whole spring of work.

Python. I had never written a line of it. Forty years after the Commodore, I was back to not knowing what I was doing — except this time I knew exactly what the thing was supposed to become, which somehow made it worse. Every day I could see the gap between the system in my head and the empty file on the screen.

Here's the part I didn't expect: it went faster. A whole spring to build the first one, weeks to rebuild it in a language I'd never used. Not because I'd become a programmer — I hadn't. Because the second time I already knew what I was building, and that turns out to be almost the entire job. The first version was slow because I was arguing with myself about what the thing should be. The second was just work.

It was a stranger kind of work, though. I wasn't building a system that watched Iran anymore. I was building one that didn't know what it was watching, and had to be told. Sources, keywords, the terms it uses to read a market — all of it had to become something the machine assembles for itself, for a subject nobody has described to it in advance.

Then one morning in July I woke up at five, like always, and the report was already on my phone.

I hadn't pressed anything. Nobody had. It had gone out to collect, read, compare, and write while I was asleep, and it would do it again the next day whether I was there or not. I sat on the edge of the bed and read my own report like a stranger had sent it to me. The first time it worked was relief. This time was different: it worked without me.

Now anyone can open a conversation with a bot and ask for a subject. Not one subject — as many as they want, each one running on its own, with its own sources and its own track record, none of them aware of the others. You say what you want to watch; the system goes looking for the voices worth listening to, checks which ones it can actually reach, and comes back with a proposal you approve or send back. I tested it on Taiwan and semiconductors — a world I know nothing about — and watched it build itself a set of sources and start counting lead times, while the Iran reports kept arriving every morning as if nothing had happened. Same machine, different war, at the same time.

Then I went hunting for the cost.

Not glamorous work. Weeks of it. Finding out which questions genuinely needed the most expensive model and which ones I'd been sending there out of laziness. Stopping the system from paying twice for the same tweet. And the one that mattered most: when two subjects — whether they belong to the same person or to two strangers — happen to need the same piece of the world, the system buys it once and hands it to both. That single change turned every new subject from a full price into a fraction of one.

The bill per user is now a fraction of what it was in the spring, and it no longer grows in a straight line when someone new arrives. That sentence is boring to read and it's the only reason any of this can exist.

Earlier in this story I wrote that maybe this was only ever going to be my personal tool.

I don't think that anymore.

And if you've read this far, let me say the part I couldn't say a month ago: Narrative Edge works. Not as a demo, not as a thing I babysit. It wakes up before I do, it runs on more than one subject, for more than one person, and it costs what it should.

So if you have a subject you keep refreshing at odd hours, waiting for the news to catch up — Narrative Edge is for you. Start here on Telegram, ask for access, tell the bot what you want to watch, and your first report can be on your phone tomorrow morning.

And if you build or back things like this: what stands in the way now isn't the system, it's me. One person, one laptop, one language he learned by accident. The machine doesn't care whether it follows one subject or a thousand — I do. That's the only ceiling left, and it's the kind that money and people remove. Hit me up here and let's see if we can grow this further together.

Narrative Edge is the first venture of Mount Dragon.