Notes

The last milestone of my bare-metal STM32 stack was an I2C driver reading a real sensor. Two bugs taught me more than the other four milestones combined — including one that only 'worked' by accident.

I wrote a bare-metal STM32 driver stack from the registers up — no HAL, no vendor code. The hard part wasn't the drivers. It was proving each one actually worked instead of assuming it did.

The hardest part of my caching proxy wasn't building it four ways. It was finding out my measurements were wrong — and using a queuing-theory law to prove it.

Part 4: I built an epoll event loop like nginx uses, benchmarked all four architectures head-to-head, and found that the server with the highest throughput was also the worst one to actually run.

Part 3: I built the fix I'd diagnosed, proved it worked, then swept worker counts and found the tuning curve. The lesson isn't 'more threads.'

I measured my caching proxy instead of assuming. Threads made it slower. LFU collapsed. Both taught me more than the 29x number did.

A custom domain, link previews, and SEO — the invisible work that makes a site feel real.

In one session I took a caching reverse proxy from a sealed Raspberry Pi box to forwarding live traffic in C. Next up: the numbers.

I almost left my best experiences off my portfolio because they weren't 'engineering enough.' I was wrong.

I set out to add a few things to my portfolio. What I actually learned was how to debug.

Why I'm keeping a build log alongside the projects — and what I plan to write about.