ESEN

Week 1 Summary: From concept to PowerFactory in the cloud

•4 min read

Days 1-7 / 60

First week of the challenge completed. 11.7% of the journey done. Here's what happened.

The starting point

A week ago I started this public experiment: building an AI agent capable of preparing ECAP studies (Protection Coordination and Settings) for the Chilean electrical system. 60 days. No excuses.

Why ECAP? Because it's the most complex study in the connection process. If the agent can handle this, it can handle the rest.

The architecture I chose

I decided to use the ReAct pattern (Reasoning + Acting). It's nothing sophisticated, it's the current standard for agents. The key difference with a regular LLM: an agent executes, not just generates text.

The tools I need to build:

  1. DIgSILENT PowerFactory — The electrical simulation software
  2. Coordinator's PGP — 6,000 projects with real studies for the agent to learn from
  3. Infotécnica — The technical database of the electrical system
  4. Report writing — Iterative generation, not monolithic

The drama of the week: DigSILENT PowerFactory

I started with the most critical and most difficult tool. Mistake.

Lost day #1: All my computers are Macs with M processors (ARM). PowerFactory only runs on x64/x86 architecture. I tried virtualizing Windows. It didn't work.

Lost day #2: I migrated to Azure because it was Windows. I hate Azure. I gave up after hours configuring a VM.

Final solution: Google Cloud. A VM with Windows, 4 vCPUs, 16GB RAM. Running in minutes.

What's running

At day 7, I have:

  • PowerFactory running in the cloud
  • A FastAPI backend exposing the API
  • MCP (Model Context Protocol) connected to test directly with Claude
  • Working tools: load project, power flow, short circuit

I can now ask Claude "run a power flow on this project" and it executes it. However, since I've only worked on the tool and nothing on the agent, it feels rough, dumb, it returns superfluous information with little utility—it literally runs the simulations and returns them to me. I feel like something's missing, I hope that by improving the agent the tool will feel better.

What's next

This week was about infrastructure. I'll continue on infra for a couple more days, I think I'm at 30-40% of what I can do with DigSILENT so I'll finish polishing this before moving to the next tool.

Pending:

  • Benchmark different machine configurations to improve simulation time.
  • Queue system for multiple simulations.
←Back to blog