RUNTIME

Time in Amaravati:--:--

Say hello

Engineering Clinics, VIT-AP

Smart Vision Screening Device

Edge-AI eye screening

A low-cost, portable eye-screening device. It photographs a patient's eye, runs a machine-learning model on the device itself, and sends the result to a web dashboard where a doctor reviews it and issues a prescription.

Prototype readyRough cutRuntime: 1 min read

The code is in a private repo. Happy to walk you through it: ask me.

Context
Engineering Clinics, VIT-AP
Guide
Dr. Vasavi Sanikommu
Role
Web dashboard developer, ESP32 firmware developer, and ML model integration
Team
6 people
When
July 2026 – September 2026

The problem

Many eye conditions are treatable if caught early, but screening needs a specialist and equipment that rural and camp settings don't have. This device is meant to be carried to the patient, used by a health worker, and to flag who needs to see an ophthalmologist.

How it works

  1. 01XIAO ESP32S3 SensePhotographs the eye and runs YOLOv8-Nano on the device
  2. 02Cloud FirestoreScores plus the image, uploaded after analysis
  3. 03Web dashboardA doctor reviews the result and issues a prescription

Key decision

Run the disease-detection model on the device itself (YOLOv8-Nano, INT8-quantised to ~1.8 MB, in TensorFlow Lite on the XIAO ESP32S3 Sense) instead of streaming images to a server or cloud API. The device uploads only the result, scores plus the image, to Firestore afterwards.

Why

The device is meant for rural screening camps, where a reliable connection can't be assumed. If analysis depended on the network, a weak signal would mean no result at all; doing it on-board means screening still works, and the upload can catch up later. It also keeps the patient's eye image away from third-party processing, which matters for medical data, and removes the per-call cost and latency of a cloud vision API. That drove the hardware change from the ESP32-CAM to the XIAO ESP32S3 Sense, whose vector AI instructions and 8 MB of PSRAM can hold the model and frame buffer.

Models

  • YOLOv8-NanoINT8-quantised to ~1.8 MB, TensorFlow Lite for Microcontrollers on the XIAO ESP32S3 Sense

Stack

  • C++
  • Arduino
  • TensorFlow Lite for Microcontrollers
  • YOLOv8-Nano
  • XIAO ESP32S3 Sense
  • HTML
  • CSS
  • JavaScript
  • Firebase
  • Node.js
  • Inline SVG charts

The dashboard

  • HTML5, CSS3 and vanilla JavaScript (ES2020): no framework and no build step.
  • Firebase JS SDK 10.14.1 (compat) with Cloud Firestore, Firebase Authentication, Firestore Security Rules and Firebase Hosting.
  • Charts drawn with inline SVG, and prescriptions printed with the browser’s own printing.
  • A custom JavaScript dictionary for English, Hindi and Telugu.
  • Node.js with firebase-admin for setup.

The firmware

C++ on the Arduino framework for the ESP32, with TensorFlow Lite for Microcontrollers running the model on the board.