This was built for the SDx ai hackathon. It was a 36 hour hackathon.
Our idea was: can we train a CNN to reconstruct LCR circuits based on probing the inputs and outputs?
This goes in line with the Thevenin/Norton rules that prove that any linear two-terminal network reduces to one source plus one impedance.
Our results were... good? Naturally, the presentation mattered more than the technical achievements. We went through 3 rounds of judges and we ended up winning the qualcomm bounty aka a couple hundred bucks.
This was supposed to be our hardware in the loop setup:
We published our results on github in case you are interested.
We also got to spend time with qualcomm ai execs, and learned that their elevators can cancel a floor if you double tap the buttons.
