60 days for an AI to sign a technical report

Day 1 / 60

I missed writing.

After 60 days documenting how I built Nelson, I took a break. But two weeks after launch, something keeps bugging me: Nelson still hasn't passed a real test.

I've had several conversations with consulting firms and interested companies. The product catches their attention. But one question keeps coming up: Has Nelson passed a study before the National Electrical Coordinator (CEN)?

No. Not yet.

And they're absolutely right to ask.

An electrical study in Chile isn't worth the report it generates. It's worth it if the CEN approves it. Without that, Nelson is an interesting tool in demos, not something a consulting firm can use in a real project.

That's what I intend to solve in the next 60 days.

The challenge

In the first series I built Nelson from scratch: architecture, tools, parallelization on Google Cloud, integration with Infotécnica, CNE and Open Access. By day 60, Nelson could run autonomous studies of up to 60 minutes and generate reports of over 20 pages.

But "it works in my tests" isn't enough.

This series has a concrete objective: for Nelson to submit a report to the CEN and have it approved.

If it succeeds, I have something real that I believe could revolutionize the market.

What I need to solve

I'll be honest about the three problems I have today.

1. Get a real case

I need a client who needs to submit a study to the CEN. It could be a consulting firm working on an active project, or a direct client who needs to connect to the SEN.

This is the hardest part. It's not a technical problem, it's a business problem. I need to sell, convince and probably do the first study at very low cost or free. I have two consulting firms in active conversations. I hope to move fast.

2. Get a real DIgSILENT license

The license I have is from university, maximum 50 buses. The SEN has over 2,500. For a real study I need a full license. The cost is significant and I'm still figuring out what to do about this.

3. Improve Nelson for real reports

So far Nelson has run small studies. Making a report that meets CEN standards — format, technical criteria, levels of detail — is another thing entirely. I know Nelson can get there, but it needs work.

The ECAP is another huge challenge as well, it's not like the power flow or short-circuit study and in a large system I'm concerned about what Nelson can do, but I insist, I know it can get there.

Why I think it's possible

What gives me the most confidence is this:

Agent temporal autonomy graph

Source: METR, March 2025 — today's LLMs can complete software tasks of up to 10 hours autonomously

Today the best LLMs can do software tasks for 10 hours autonomously. Nelson has reached 60 minutes of running. I believe in 60 days I can take it to multi-hour studies. And in two months the models will keep improving, pushing the limit upward.

Plus the HLE benchmark keeps advancing, ARC AGI 3 also claims to be completely agentic. Agentica says it has completed ARC AGI 3 on the games they have so far, which are 6.

So clearly technology is on my side. I need to bring the human side to solve what's left.

The goal

Productivity experts say goals need to be clear and measurable. This one is:

If in 60 days Nelson submits a study that the CEN approves, the challenge is complete.

I believe it's possible. But in 60 days I'll know for certain.