01On-device ML

MoodScope

Health Hacks @ USF 2025 · 2025

Real-time facial-emotion recognition, quantized for the edge

~33 FPSon-device on a Raspberry Pi 4

Reads emotion from a face in real time on a Raspberry Pi 4, using INT8 quantization to run entirely on the edge so nothing ever leaves the device.

An on-device facial-emotion detector built for edge hardware: INT8-quantized TensorFlow Lite models hit ~30 ms per frame (~33 FPS) on a Raspberry Pi 4, and threaded dual detection backends cut frame stalls by ~45%. A local Llama 3.2 turns the readings into plain-language insight without any data leaving the device. Built at Health Hacks @ USF 2025.

Stack
  • TensorFlow Lite (INT8)
  • Llama 3.2
  • OpenCV
  • Raspberry Pi 4
  • Python
01

How it's built

SPECIFICATION
Device
Raspberry Pi 4 (edge)
Models
TensorFlow Lite, INT8 quantized
Throughput
~30 ms/frame (~33 FPS)
Latency
~45% fewer frame stalls (threaded dual backends)
Insight
local Llama 3.2, on-device
Vision
OpenCV
MY ROLE
  • 01Quantized the detection models to INT8 TFLite for real-time inference on a Pi 4.
  • 02Built threaded dual detection backends, cutting frame stalls ~45%.
  • 03Integrated a local Llama 3.2 for private, on-device insight.
Saksham Srivastava | Full-stack Developer