Published:
The original robotic writer picks up letter blocks and places them one at a time. It is a pick-and-place operation — move to the letter, grasp, move to the writing line, place, return. Each letter is a discrete event with its own approach, grasp, and retract cycle. For printed text, this works. For cursive, it is fundamentally wrong. Cursive means the pen never leaves the paper between letters. The trajectory is one continuous, flowing curve.
Published:
Standard SOFI needs hundreds of frames of blinking emitters to beat the diffraction limit. You acquire a long temporal stack, compute cumulants across time, and the nonlinear statistics sharpen the point spread function beyond what any single exposure can resolve. It works beautifully — but it demands fluorophores that blink stochastically, and it demands patience while you collect enough frames for the statistics to converge.
Published:
Rewrote the industry portfolio entries today. Realized that the way I described these projects – starting with XGBoost hyperparameters and Kedro pipelines – was the view from the developer’s chair. But I haven’t sat in that chair for a while now. The +100 TPH isn’t interesting because of the model architecture; it’s interesting because of what it means for the operation’s annual throughput. The 15% alert reduction matters because a mine that violates environmental regulations can be shut down. Added business impact sections, KPI tables, strategic context. The technical details are still there – they just come after the “why it matters” part. It’s a small reordering that changes the entire story. The same work, viewed through a different lens, tells you something different about who did it and why. Feels more honest about where I actually spend my time these days.
Published: