Ark of Noahledge Offline · Solar · AI Knowledge node · Open source forever — democratizing knowledge where it's needed most
About

I put AI into production
for a living.

Then, in the space of seven weeks, I watched two earthquakes take the internet away from millions of people — and the second one took it away from the city I grew up in.

Who is building this

Juan Hurtado

I lead business intelligence, data and analytics at a Fortune 500 industrial manufacturer — an S&P 500 company — where my teams design, deploy and govern AI agents that thousands of people depend on to do their jobs. The same remit covers the analytics infrastructure underneath them, the data science practice, and the data strategy that decides what gets built at all. Retrieval over messy real-world documents, model selection, evaluation, the governance that decides when a system is trustworthy enough to put in front of someone: that is my working week.

Earlier in my career I worked in an embedded hardware and software lab at a large semiconductor company, where the machine in front of you was the product. A mistake there cost a board revision rather than a redeploy. That is where I learned what a power budget actually is, why a part you cannot source is worse than a part that is merely slow, and how much of a system's reliability is settled before anybody writes a line of code.

Almost everything I know about this subject came from systems that had to keep working on a factory floor at three in the morning — not from demos. This project sits exactly where those two halves meet: a hardware build inside a strict power envelope, carrying a retrieval system that has to find the right four paragraphs in a terabyte. That is the whole of what I know pointed at one problem — and the reason I am giving it away rather than keeping it.

What that background is worth here

The hard part of this project is not running a model offline. Anyone can do that in an afternoon. The hard part is retrieval quality — whether the right four paragraphs out of a terabyte actually reach the model — and knowing when the answer should not be trusted. Both are things I do at scale for a living, on corpora that are messy in exactly the way real archives are messy.

And what it isn't worth

Nothing in my background makes me a seismologist, a doctor or a disaster-response professional. Where this project touches those fields it defers to published references and says so, which is why the whole design is built around citing a source rather than sounding confident.

The reason

Two earthquakes, seven weeks apart.

This project existed as an idea for a long time before it had any urgency behind it. Then it acquired some.

24 June 2026

Venezuela

M7.2 & M7.5
39 seconds apart

Two earthquakes struck Yaracuy thirty-nine seconds apart — the strongest Venezuela had recorded since 1900. More than 1,900 people were killed and hospitals across La Guaira and Caracas were damaged or running at partial capacity.

The detail I could not stop turning over was one that happened 1.8 kilometres offshore. The South American Crossing submarine cable was severed. Roughly half of the country's international bandwidth ran through it, and repairing it took a month.

Nothing about the knowledge itself had changed. Every medical reference still existed. Every repair manual, every textbook, every article still existed. What broke was the path to them — one cable, in one place — and a country's worth of schools, clinics and factories was on the wrong side of it.

The communities that stayed dark longest were the ones already living on weak, intermittent connections. Some access was eventually restored by volunteers running more than two hundred hotspots off satellite terminals, batteries and solar panels — which is a fairly exact description of what a fallback layer for this ought to look like.

1,900+Lives lost
~50%National bandwidth cut
1 monthTo repair one cable
1.8 kmOffshore, where it broke

I started building this in the evenings. Nights, weekends, whatever the week had left over.

10 August 2026

Colombia

M7.4
07:34 local

Seven weeks later, at half past seven in the morning, a magnitude 7.4 earthquake struck at San José del Palmar, in Chocó. Fifteen of Colombia's thirty-two departments were affected. Among the hardest hit were Quibdó, Cali, Manizales, Armenia — and Pereira, which is where I am from.

The damage to the things this project exists to serve is almost too on the nose: 2,838 schools and 285 health centres damaged, along with 407 roads and 44 bridges — the connections of every kind that let a place ask for help and receive it.

And then Chocó. The epicentre landed on the poorest department in the country, one reachable mostly by river, where officials have been candid that the full picture is still not understood. That is the part I keep coming back to. The hardest place to reach after an earthquake is the hardest place to reach on an ordinary Tuesday. It is last to be connected, last to be assessed, last to be helped — and first to be cut off.

The first earthquake was something I read about with a professional interest in the failure mode. This one is home.

289Lives lost
15 / 32Departments affected
2,838Schools damaged
285Health centres damaged

So I stopped treating this as something I get to when the week allows.

Figures as reported on 21 August 2026, drawn from published reporting on both events. The Colombian totals are still moving — 4,187 injured, 143 officially unaccounted for, and a civilian missing-persons database listing many times that. They are cited here because precision is the least this deserves, not because the accounting is finished.

The conclusion

What I actually took from it.

First, the honest limit. An offline knowledge node does not stop a building from falling down. It does not pull anyone out of rubble, and it would not have saved a single life in either of those weeks. Nothing I am able to build does that, and I would rather say so plainly than let this be mistaken for disaster relief.

What it does address is the long afterward — and the ordinary before. The clinic still standing that needs a treatment protocol with no signal. The school that keeps teaching while the tower is down. The engineer who needs the manual for a pump that arrived without one. The months of reconstruction after the cameras leave.

And more than any of those: the places where "cut off" is not an emergency but the permanent condition. Chocó did not lose a good connection on the tenth of August. It never had one to lose. For a community like that, a box that already holds the medical references, the textbooks, the repair guides and the maps is not disaster preparedness. It is just the library finally arriving.

Three design consequences follow, and every decision on the Architecture page comes from one of them:

It has to already be there

Knowledge delivered after the event arrives over the infrastructure the event destroyed. The only version that works is the one sitting on the ground beforehand — which means the whole thing has to be cheap enough, and dull enough, to install somewhere nothing is currently happening.

It has to make its own power

The grid goes when the network goes. A node that needs mains electricity has simply moved the dependency rather than removed it. Hence the watts-first design and a solar budget sized before the compute.

It has to survive me

A system only its builder can operate is not a backup. Printed recovery instructions, vendored dependencies, three copies, hashes on every file, and a named second person who runs the drill — because the day it matters, I might be somewhere else entirely.

One cable, in one place, and a country of schools and clinics was on the wrong side of it.
Get in touch

Who I'd like to hear from.

Schools and clinics

Particularly in Chocó, Risaralda, Valle del Cauca, Caldas and Quindío. If you work somewhere that spent August without a connection and can tell me what you actually needed to look up, that is worth more to this design than any amount of theorizing.

NGOs and funders

The hardware is a one-time cost and the knowledge is free. A single node serves a whole building over local WiFi. If you deploy in places where connectivity is unreliable and want to talk about what a pilot looks like, I would like that conversation.

People who want to build one

Every model here is openly licensed and every corpus is freely available. Nothing about this is proprietary and nothing needs my permission. If you would rather build your own than wait for anyone, good — tell me what breaks, so the instructions get better.

Contact

The fastest way to reach me is LinkedIn.