More capacity for your operation, without losing control.

For power grids, plants and payments. AI proposes, your rules decide what runs and a person signs off on what matters.

When AI acts on your operation, a mistake has no undo.

A breaker opens. An arm closes. A payment goes out. Before a model touches anything real, we define what happens and what doesn’t.

Four layers between the model and your equipment.

The model proposes. Everything else decides what runs. Already have AI connected to your operation? We review it against these four layers.

Reasons about what it sees and proposes. It never executes on its own, not even the obvious.

It only sees what we describe of your operation and only does what is enabled. Anything off the list does not exist for it.

Check every input and every proposal against hard limits: ranges, schedules, plant states. Whatever falls outside is held.

What matters is signed by a person before it runs. Every step is written down, with time and owner.

scroll to open
Franco Pérez

Who’s behind it

Franco Pérez, Computer Engineer. Seven years building systems where a command can’t be undone.

I talk to RTUs over IEC 104 and design systems that can’t be forged. Crisol joins the commanding protocol with the security that authorizes it.

  1. With Digicom, an IEC 101/104 module that reads and commands inside a substation SCADA, 49 requirements traced. Before that, the SCADA for Tucumán’s grid.

  2. In automotive manufacturing, real-time plant-floor data across 8 lines and 32 machines. I led the team that built it.

  3. Payment routing that raised approval rates in production. Today, with Littlebird, a prediction market with custody and on-chain settlement.

  4. Zero-knowledge tooling with Aztec and Noir, awarded at the Solana Privacy Hack. An Ekoparty talk on identity-less authorization.

How we work

We look for quality in the process, from the first conversation.

We start with your operation: who decides and where an error is expensive. The specification comes before the code: what can’t break and what gets tested.

AI agents build against that specification and we review every delivery. In operation we observe the system and what we learn goes back into the specification.

WHAT WE LEARN GOES BACKConversationwho decides, where an error costsSpecificationwhere our judgment goesAgents buildagainst the specificationWe review every deliverynothing ships unreviewedObserved operationevery action, traced

Your equipment in SCADA

RTUs and field devices of any brand, in their protocol: IEC 104, DNP3, Modbus. We read first, without touching the network. Commands come in proven.

IEC 104 · 61850 · DNP3 · Modbus

Your AI, within limits

If you already have agents or robots on your systems, we bound them: closed permissions, a person approves what matters, everything logged.

Agents · robotics · closed loop

The Crisol Criterion.

Criteria for evaluating an AI system before authorizing it to act on your operation.

Open it in your agent: it interviews you about the system you are being offered and returns a one-page report.

Read https://crisol.studio/criterio-en.md and follow the "For the agent" section: interview me about the AI system I am being offered and return the report.

Read the document

Questions

What does Crisol Studio do?
We build and maintain systems that cannot fail: SCADA and substations, read-only plant data, and payments, with AI that proposes and a person who approves what matters.
How does a project with Crisol start?
With a fixed-price diagnostic. You leave with the map of your risk and what to build, whether you continue with us or not.
Does the AI touch the control network?
Not at first. We start with passive listening and read-only access; each level of authority is earned with evidence and with human approval.

Let’s start with the process that has the least room for error.

Book a meeting: walk us through that process and leave knowing what to build first.