What this is. An amplifier is only linear up to a point. Push the input past that knee and the output no longer follows it, and the error shows up as harmonics, mostly the third. Predistortion means bending the input the opposite way first, so the amplifier's own curve straightens it back out. The interesting part is that the right bend changes as the amplifier warms up, so the correction has to be measured and re-fit rather than set once.

A diode-clipped op amp stage distorts once it passes its compression knee, and the dominant term is cubic. I am designing a polynomial predistortion stage that applies the inverse of that curve ahead of the amplifier so the chain stays linear further into the knee. The model lives in LTspice; the coefficients are fit in Python from DC sweep data rather than derived by hand.

Fixed coefficients drift with temperature and load, so the second half is an SDR feedback loop that captures the amplifier's actual output and re-fits the inverse with a learned predistorter. This entry gets numbers when the loop closes.

What this is. A PWM signal sets how hard a motor is driven. Most controllers check for faults in firmware, which is fine until the firmware itself is the thing that has gone wrong. Here the fault logic is in the FPGA fabric, one clock cycle from the output, and once it latches there is no software path to unlatch it. The point was to build the safety envelope in a place a bug cannot reach.

When a fault asserts, the output drops to a safe state and stays there until an operator resets it, so a transient glitch cannot quietly put the drive back in motion. A separate configuration-validation module rejects out-of-range parameters at load time rather than letting bad settings reach the driver stage. Carried through the complete flow, from simulation to a physical board.

What this is. A processor pipeline works on several instructions at once, one per stage. That breaks when an instruction needs a value the previous one has not finished producing, called a hazard. Forwarding hands the value across early; stalling waits. I wrote the model in C so I could watch every stage every cycle, and wrote the compiler that feeds it so the test programs were real C, not hand-picked assembly.

Most of the work was hazard debugging, tracing execution cycle by cycle in gdb and comparing register state by hand, because a wrong value three instructions later usually originates several cycles earlier.

What this is. A camera on a roadside unit watches for animals and warns drivers. It has to run on a small board with no network, so the detection model is shrunk from floating point to 8-bit integers and moved onto a dedicated accelerator. Shrinking a model loses accuracy unless it is done carefully, which is what the calibration set and the quantization-aware fine-tuning are for. The rest of the work was making it survive real hardware: a service that comes back on its own after a power cut, and tests that run without the board attached.

The Hailo compiler refused the model. I traced it to an incompatible automatic postprocess-extraction step and rebuilt the export path with explicit end-node parsing and a custom NMS configuration rather than working around it. The camera calibration replaced a placeholder focal length that every prior distance measurement had silently depended on.

Bugs fixed on the real unit: a SIGTERM handler bypassed by a vendored library, a non-daemon thread that hung shutdown, a race between the alert thread and hardware cleanup (confirmed with a 20-trial stress test, fixed with a lock and idempotent teardown), a flaky I2C connector, and an apparent startup hang that was stdout block-buffering.

What this is. A convolutional network that looks at a photo of a leaf and names the disease. The number that matters is test accuracy on images the model has never seen. Getting it up meant reading the confusion matrix to see which diseases the model mixed up and why, then changing the network in response, rather than re-running training with different knobs and hoping.

What this is. A publishing platform for verified student journalists, built and run by one person. The part I would point a reviewer at is not the feature list but the security review: before launch I went through sixteen ways the site could be abused and found two real ones. Fixing the instance is easy; the rate limiter exists because I wanted the whole category closed.

The access-control flaw let a signed-in user reach another user's records through a predictable API route. An LLM API is integrated as a production feature, not a demo. When something breaks I read the logs, reproduce it, and write the test that should have caught it.