Years ago, I gave up on a thermometer project. The synthetic thermistor model I was using was inaccurate — off by up to 2°C in places — and I set the electronics aside to go back to programming, mostly to shake off the rust after years away from it.
That's how nnetwork was born — later renamed Hasaki. Just an exercise in understanding how neural networks work, nothing more.
Then one day I looked at that old thermometer again and thought: what if the formula really is the problem? Why not ask Hasaki to help me find the right correlation?
Later, while building a smoke detector — a device where every millisecond counts — I needed something faster and lighter than neural inference. That's how kigu-quant was born: a tool for generating static C lookup tables so tiny chips can process math like sigmoid without an FPU. Hasaki's exported header relied on expf(); kigu-quant replaced it with a lookup table. No floating-point exponential, no FPU dependency — just a lookup and done. More on that here.
With the thermometer, the lesson was different. We eventually found that a well-fitted formula beats inference — especially once we measured the thermistor's real Beta and Ro instead of trusting synthetic values. The final solution: a 5th-degree polynomial fit, with kigu-quant handling the lookup table for the exponential term. Hasaki wasn't wrong to try. It just wasn't the right tool for a problem that already had a closed-form answer.
Seeing kigu-quant solve a real problem got me thinking: what other friction points do embedded engineers deal with constantly? UART timing calculations kept coming up. That became kigu-uart.
There's also csv_prep — currently closed, being tested against real datasets for the next major version of Hasaki. It'll ship free, bundled with that release.
Rounding out the toolbench: a bidirectional resistor color-code calculator, and QUAD7SHIFT, a driver library for 74HC595-based seven-segment displays.
I never imagined an exercise in understanding neural networks would turn into a passion for building tools that are small in size but big in impact — tools that help others, not just me.
And I never thought I'd actually publish them, instead of shelving them the way I did for years with the microcontrollers I'm only now getting around to rescuing.
Top comments (0)