Neuropilot

I was reading about non-invasive brain stimulation and wanted to know what happens inside a network of neurons when you stimulate it. Simulating it is how you find that out, so I installed The Virtual Brain, Brian2 and NEST.

Then I spent my time on installation and parameters instead of on the question. Writing a correct simulation is a skill of its own, separate from having the idea, and it is the one that eats the week.

So I built an agent to take that part over.

Two versions

v1 configured predefined Brian2 simulations from a question asked in ordinary language, searching the documentation and the literature to do it, and gave the results back synthesised instead of as raw output. Its limit was autonomy. The simulations sat in fixed n8n workflows, so the agent could choose and parameterise but never shape a simulation around a question the workflow hadn’t anticipated. Research questions rarely fit a menu.

v2 removed the menu. I fine-tuned a model on the Brian2 documentation and gave it a sandbox to execute in, so it writes and runs the simulation script itself.

Showing it to people who would know

In March 2025 I posted it to the mailing list of the Organization for Computational Neurosciences and asked for feedback.

It came back from Chile, the UK, the US, Germany and Hungary. PhD students and postdocs, and two professors who had built their own neurosimulation teaching tools decades earlier.

Almost none of it was flattering.

A researcher in Birmingham tried to replicate his own published experiments. He asked what happens to cerebellar activity during eye-hand coordination, watched the agent run all the way through to report generation, and got an error after five minutes. In German. A PhD student in Valparaíso handed it a 6x6 functional connectivity matrix and asked for the mean. It said 0.4. The answer was 0.6, and he wrote back that a tool like this is worth nothing if it can’t handle a number reliably. A postdoc at UCSF couldn’t find any way to switch the interface out of German. Several people just wanted to know which models were available to them, asked, and got an error.

Why I stopped

For a boring reason. The frontend was on Vercel and the backend on Google Cloud, I never got the asynchronous deployment between the two working properly, and at some point in May 2025 I was frustrated enough to put it down.

It is still the project I keep coming back to. Getting agents to do science that holds up is the question I find most interesting, and this was my first attempt at it.