Running Neural Networks Inside the Web Browser? Only if You Know How!

Client Overview

ShareTheBoard is an educational technology platform that transforms handwritten whiteboard content into clear, accessible digital visuals, using intelligent transcription and real‑time vectorization to support learning in any modality.

Challenges

High server‑side GPU costs made traditional real‑time neural network inference impractical.
Requirement to run neural network inference client‑side within standard web browsers.
Existing technologies (WebAssembly, WebGL/WebGPU) struggled with models using high res images as input and output.
Early implementations caused the browser to freeze for seconds, making the app unusable.
Framework‑level tools like TensorFlow.js and ONNX.js created UI blocking or were too slow.
WebGPU worked only on high‑end devices and degraded performance on older hardware.

Solutions

Developed a proprietary JavaScript library using WebGL + GLSL shaders.
Broke model execution into smaller layer‑level steps to allow browser rendering simultaneous with inference.

Work we’ve delivered

Custom Browser Neural‑Inference Engine

Proprietary WebGL/GLSL library allowing real‑time neural inference without freezing the UI

WebGL

Browser inference

Neural network optimization

Layer‑by‑Layer Execution Pipeline

Divided neural network into granules for controlled GPU use and smooth rendering

Performance tuning

Browser GPU control

Extensive Device & Browser QA

Tested across hardware/software configurations to ensure consistent performance

Cross‑device testing

Edge compatibility

No items found.

Tech Stack

Frontend

JavaScript

WebGL

GLSL

shader programming

Custom R&D solution

Backend

JavaScript

WebGL

GLSL

shader programming

Custom R&D solution

Platform

JavaScript

WebGL

GLSL

shader programming

Custom R&D solution

DevOps

JavaScript

WebGL

GLSL

shader programming

Custom R&D solution

Design

JavaScript

WebGL

GLSL

shader programming

Custom R&D solution

Other

JavaScript

WebGL

GLSL

shader programming

Custom R&D solution

Results

Successfully released the convolutional neural network with FullHD resolution for input and output inside a fully functional web application.
Real‑time content detection and digitization without browser freezes.
Smooth performance across diverse hardware configurations—from older laptops to modern devices.
Enabled ShareTheBoard’s commercialization by reducing hosting costs and increasing accessibility.

Conclusion

We took ShareTheBoard’s powerful neural network from server‑side prototype to polished, browser‑native solution by building a custom inference library with WebGL and shader‑based layer execution. This approach addressed performance bottlenecks, eliminated UI freezing, and ensured compatibility across devices. The result: real‑time handwritten content detection fully embedded in a web app, delighting users and enabling scalability.

Meet the minds behind the project

We're the brains and the heart behind the code. A quirky bunch of passionate pros who love turning ideas into reality. Here, every project is a team sport, and we’re all about building software – and relationships – that last, one line of code at a time.