Skip to the content
Krizaka Blog

01

Why I am building Krizaka in Montréal

Oct 16, 20265 min read

What I faced, A claim without a source; What we built, a run that keeps its id, its version and its rules

A watercolor of Montréal rooftops at dawn under a pink sky, a small bird perched on a chimney

This week our demo videos were generated on an NVIDIA H100 rented from RunPod. In the rental log, the pod's region is called CA-MTL-1. Montréal. The machine that made our pictures was running in my own city.

That was not the plan. I wanted a 48 GB GPU at DigitalOcean, in Toronto, where our databases already live. There was never one in stock in that region, the only one our private network reaches. So we took the H100 in Montréal. Looking at that region name, it struck me as a good place to start explaining what we do.

First, who is writing

My name is Oussama, and I build Krizaka in Montréal. I studied mathematics and computer science. My first brush with what we now call AI was in 2011, in a research lab (about the author). I founded Krizaka here, and I build it mostly on my own. A large part of the code is written with AI agents that I direct, review and correct. That matters for what follows: when you work with machines that write, you quickly get picky about who did what, and where.

Sovereign AI, in one sentence

The word gets used for everything. Here it means one precise thing: a customer can know where their document was processed, which model read it, and who can see the result.

Think of a notary. You hand over a deed. You know which office keeps it, who signed it, and there is a register. If someone disputes it, you open the register. Sovereign AI is the same thing for a document that a machine has read.

In practice, every analysis is a run of Orazaka, the engine I'm building. The run keeps its id, the version of the processing that ran, and the rules it applied. The model it used is a row in a database, not a name buried in the code. A product built on Orazaka never calls a model on its own: it goes through the engine, which keeps the record.

None of this makes data private by magic, or the model any smarter. It lets you check. A promise you cannot check is marketing.

A promise

  • Where? “In the cloud”
  • Which model? A name buried in the code
  • Why this answer? An opinion, with no source
  • Disputed? Nothing to open

A run of Orazaka

  • Where? The machine's region, such as CA-MTL-1, Montréal
  • Which model? A row in a database
  • Why? The rule and the exact passage, or silence
  • Disputed? The run's id, its version and its rules: the register
The three questions to ask any AI tool, answered two ways: by a promise, and by a run that keeps its record.
A winter night in Montréal, made with Orazaka: five seconds, generated by a hosted model. · 0.5 MB
How this was made
Model
Wan 2.2 T2V A14B · Apache-2.0
Engine
hosted: DigitalOcean serverless inference (steps and guidance not exposed)
Machine
hosted by the provider (not our machine)
Through
orazaka video --model wan2-2-t2v-a14b, routed by Orazaka
Prompt
Winter night in Montréal, slow tracking shot along a snowy street of the Plateau with outside spiral staircases, warm windows glowing, snowflakes in streetlight, steam from a vent, cinematic 35mm, shallow depth of field. No text, no lettering, no logos.
Seed
20261011
Size
832×480 · 16 fps, 5.06 s
Time
132 s, as recorded by the run
Cost
US$0.60 (the provider's price)
Sovkora run
0515ce61-0e91-4fe9-a0bf-2ca2838f74e3
Sound
music bed made by Orazaka (Studio video-sound, ACE-Step 1.5, MIT), -14 LUFS, AAC

Why here

The law I read first is Quebec's: the Civil Code, and Law 25 on personal information. A Quebec lease, an employment contract, an invoice from a local firm: those are the documents our first tools have to understand. And the people who will have to trust the result live here too. I'd rather build a tool that reads a Quebec lease knowing I'll have to live with its answer myself.

Then there's language. Every Krizaka site ships in French and English on the same day. So does this blog: I write in French first, then make the English version.

What it forces me to do

Choosing to be checkable has a price, and I'd rather say so.

A result without a source doesn't show up. When a tool flags a problem in a document, it cites the rule and the exact passage. If it can't cite, it stays quiet. That is frustrating at first, because a model always has an opinion. But an opinion without a source is worth nothing in front of a client.

I keep the infrastructure small. No service I can't explain, no big machine left on for nothing. I rent a GPU when there is work and switch it off when the queue is empty. I learned that rule by paying for it, and the story is in the post about costs.

I publish the code of the foundations. The interface kit (Krizaka UI) and the Spring Boot building blocks (Krizaka Java) are open source. Anyone can read what our products stand on.

And I say what isn't finished. Several products are marked "in development": sign-up is closed, and their sites say so.

  1. AnswersNo source, no resultA finding cites the rule and the exact passage, or stays quiet.
  2. MachinesA small infrastructureA GPU rented when there is work, switched off when the queue is empty.
  3. CodeOpen foundationsKrizaka UI and Krizaka Java can be read by anyone.
  4. StatusWhat isn't finished says so“In development”: sign-up closed, and the site says it.
What being checkable forces me to do: four habits, each with its price, from the answer a reader sees down to what the site says about itself.

What isn't solved

Sovereignty of the model is still the hard part. Running a good model on a machine you control costs real GPU hours and real dollars. The Montréal H100 is US$3.99 an hour. I measure all of it and publish the numbers.

A rented machine isn't an owned machine. I know where our data is, but I don't build data centres. The region of a pod is something I can tell a customer. A contract with that provider is something else.

Last, our rules for reading Quebec leases are still waiting for a lawyer's review. The reports say so in plain words.

If you're evaluating an AI tool for your documents, ask it three questions. Where is my document processed? Which model read it? Who can see the result? If the answer is vague, or nobody can show you a trace, you have your answer.

This blog follows the same rule. If a claim doesn't lead back to a measurement, a run or a line of code, I don't write it.

Chapter 01

Results

an hour of the Montréal H100

US$3.99

What isn't solved