Devlog 2 - Smart Bin

ml pipeline

• 6 modules, tensorflow only, three dependencies. cut it from 629 lines to 308, and the old one trained in pytorch then converted to tflite for the pi, so it needed both frameworks
• the split is by session, not by frame. frames from one session are near-duplicates, so a random split puts near-copies of your training data in the test set and the accuracy number means nothing
• evaluate.py prints both numbers and the gap. the frame-split one is there as the number not to believe

firmware

• 11 files. trigger, weigh, light, capture, classify, decide, tilt, log, serve
• the safety rule is enforced twice: a confidence gate at runtime, and a config check that refuses to start if the rules file would ever send an unidentified item to recycling
• weight adjusts the result too. paper over 40 g is probably wet, so it goes to trash

bom

priced everything out, then checked the parts against my own schematic. found five problems:

• the mcp3208 had nothing to measure. its only connected input is the battery divider, and there’s no 12 V source in the build, so i dropped it
• the load cell “kit” ships an hx711 module. my hat already has an hx711 on it, and those modules run at 5 V where mine is deliberately at 3.3 V. bought the bare IC instead
• camera module 3 at $36 to module 2 at $9.49. autofocus is pointless when the camera and the pan never move relative to each other

before: $283.52
now: $195.19