AI

An exploratory concept investigating AI-assisted debugging workflows for scientific and engineering simulations.

AI coding systems frequently lack persistent structural understanding of repositories. They can generate code. They struggle to reason about a codebase they didn’t write. This project built a system that gives a local LLM a structural map of any Python repository - before it touches a single line of code. The Problem Large codebases are opaque. You can read files sequentially, but you can’t reason about dependency structure, identify which modules are most central, or understand what a change to one function propagates to - not without analysis tooling. ...

Modern repositories are increasingly too large and interconnected for shallow debugging workflows. Static analysis tells you about structure. Dynamic analysis tells you what happens at runtime. Neither alone tells you why a structurally important module is failing. This project combined both. The Core Idea Traditional debugging tools answer: “What failed?” This system attempts to answer: “What structurally important subsystem is most likely responsible for failure propagation?” ...

Most AI assistants are generalists. Ask them about Butler-Volmer kinetics, the difference between SEI growth mechanisms in NMC versus LFP, or how to interpret a GITT pulse sequence - and they will give you something plausible, occasionally correct, and impossible to trust without cross-checking. That is a fundamental problem for engineering work. A tool you cannot trust at the domain boundary is a liability, not an asset. The Battery Expert AI was built to close that gap - a domain-specific assistant that understands battery electrochemistry deeply enough to be genuinely useful, and runs entirely on-premise so that proprietary cell data and OEM specifications never leave the building. ...

A Mechanical Engineer Walks Into a Neural Network I need to be honest about something. When ChatGPT exploded into public consciousness, I felt left behind. Everyone was talking about it. Using it. Marvelling at it. And I was sitting there thinking: I have no idea what is actually happening inside that thing. That bothered me more than it should have. You see, I have never liked black boxes. In my simulation work, I need to know why a model behaves a certain way. Is it a physics limitation? A numerical instability? A mistake in my boundary conditions? If I cannot explain the failure, I cannot fix it. ...

Image credit: Panoramio / Wikimedia Commons Oh, Not Again! Every few years, a word escapes the labs and gets dragged through marketing departments worldwide. “Cloud.” “Blockchain.” “Metaverse.” Each promises to change everything. Each delivers a fraction. Each fades. Now it’s “AI.” This one bothers anyone who actually cares about definitions. ...

Image credit: Wikimedia Commons The Smallest Unit of the Marvel Before transformers. Before attention. Before LLMs. There was the Artificial Neural Network. The humble building block that started it all. Back in 2013, when none of this was the talk of the town, neural networks were quietly fascinating. Not because they were powerful. They were not. Compared to today, they were weak. Slow. Limited. ...

Most AI coding tools are generation engines. You describe what you want, they produce code. This works for isolated functions, boilerplate, and anything where the context fits in a prompt. It breaks down when the problem is understanding an existing system - navigating a codebase you didn’t write, debugging a failure with no obvious error message, figuring out why something that worked in isolation fails in the larger system. These are comprehension problems, not generation problems. And comprehension requires structure, not just pattern matching. ...