Accurate determination of pH is a critical requirement for water-quality monitoring in aquaculture; yet the deployment of machine learning algorithms on 8-bit microcontrollers is severely constrained by their minute memory and computational capabilities. This work reports a proof-of-concept point-of-care device for pH-estimation based on an electrochemical pH sensor with two-stage operational amplifier-based signal-conditioning circuit: low-pass filtering and temperature compensation, a lookup-table-assisted calibration model implemented on a low-power 8-bit AVR microcontroller (ATmega328P, 32 KB flash, 2 KB SRAM, 16 MHz, no floating-point unit) as well as associated mobile application for real-time visualization and alarms. Four compact machine learning models were evaluated on the target microcontroller after aggressive weight pruning and 8-bit integer quantization: a two-hidden-layer feed-forward neural network, Bonsai framework, a Tiny1D CNN and a Linear SVM. The Linear SVM algorithm demonstrated the best performance with 98.0% accuracy, 0.08 pH mean absolute error, 0.11 pH root-mean-square error, and 0.99 coefficient of determination (R²), while requiring only 1.0 KB flash memory, 36 B SRAM and 8 ms per inference. The Tiny1D CNN model achieved 91.2% accuracy but with increased resource requirements (1.8 KB flash, 158 B SRAM, 45 ms). These results confirm that computationally intensive machine learning algorithms can be successfully deployed in aquaculture pH-sensing applications with ubiquitous 8-bit microcontrollers. The proposed approach will be further validated in field conditions in brackish-water aquaculture, with long-term drift assessment and power-consumption characterization as the follow-up studies.

pH sensing;TinyML;ATmega328P;quantization;water-quality monitoring;point-of-care device;aquaculture