Now
What I'm actually spending my time on lately — inspired by the now page idea. Updated whenever something changes, not on a schedule.
Last updated:
PhD research
Working through my PhD thesis on energy production, storage and use in vehicle propulsion — mostly literature review and simulation setup right now.
Day job at TECOSIM
FEM simulation work — crash, occupant safety and structural studies, same as always, just the current projects keep changing.
Understanding AI, not just using it
Reading and experimenting to actually understand how LLMs work under the hood, instead of only using them as tools. The little chat bubble in the corner of this site is part of that — built (and still being tuned) with an AI pair-programmer alongside me.
Turning my day job into demos
A run of engineering demos for the portfolio, each one checked against a known analytical solution before I trusted it: a modal analysis solver (natural frequencies verified against the exact closed-form beam solution), a Classical Laminate Theory calculator for composites, and a crank-angle engine cycle simulation that converges onto the ideal Otto cycle as the burn duration shrinks. The newest one races a physics-informed neural network against my own FEM solver and the closed-form solution on the same beam — and reports, out loud, that FEM still wins. The same kind of math I use professionally, but small enough to run instantly in a browser tab.
Branching out into algorithms & graphics
The portfolio's Algorithms & Graphics corner was thin, so I built a batch of visual, interactive demos: a Delaunay/Voronoi mesh generator, live contour extraction with marching squares, Fourier epicycles that redraw whatever you sketch, and a Verlet cloth you can tear. Same rule as the engineering ones — each is pinned to an external truth before I trust it: the mesh to the empty-circumcircle property and Euler's counts, the contours to machine-exactness on linear fields, the DFT to Parseval and the shift theorem, the cloth to the analytic catenary a hanging chain settles into.
Two new AI demos, verification first
The AI corner of the portfolio now has a transformer and a controller, both written from scratch and both pinned to something outside themselves before a single pixel of interface existed. The micro-GPT trains live in your browser and is judged against the bigram entropy of its own corpus — the score no model that reads only the previous character can beat. The other one runs a neural policy trained by evolutionary search against LQR on an inverted pendulum, and neither gets to win by assertion: the policy has to prove it never beats the provable optimum near upright, and LQR is measured standing the pole up from 2 hanging starts out of 30. Building the interface caught a bad check in that one — the harness had declared LQR incapable of a swing-up after watching it for only ten seconds, and the browser, left running for thirty, showed it doing exactly that.
Writing more
Trying to actually write up the demos I build instead of just shipping them quietly — the blog is where that's landing.