How to Create Your Own Neural Network?: ShareTheBoard Case Study

Building STB‑745: a lightweight, browser‑ready neural model for real‑time whiteboard content detection

Founded
Interviewed
Jerzy Nowiński
Industry
Educational Technology / Remote Learning
Favourite feature
The custom STB‑745 neural network for browser‑based segmentation

Client Overview

ShareTheBoard bridges the gap between traditional whiteboards and digital collaboration tools. It captures handwritten content in real time, clears away visual obstacles (like the person writing), and presents everything in a crisp, zoomable digital format. Unlike digital whiteboards, ShareTheBoard preserves the natural writing experience and boosts it with powerful features like automatic saving to Board Memory and AI-powered transcriptions or meeting recaps. The result? Whiteboard content that’s always legible, accessible, and ready to share, during the meeting or long after it ends. Its impact has been recognized widely: - Winner of 2025 Global Engagement Innovation Award. - Winner of District Administration’s Top EdTech Product Awards at the FETC® (Future of Education Technology Conference) - Named one of the Top 10 Distance & Remote Learning Solution Companies of 2022 by Education Technology Insights

Challenges

  • Required real-time semantic segmentation to distinguish content, background, and disruptions on the whiteboard.
  • Neural network behind it needed to be small and efficient enough to run smoothly in standard web browsers, including on older devices.
  • High-resolution image processing demanded both speed and lightweight architecture.

In short: deliver accurate real‑time AI on low-powered devices via the web.

Solutions

  • Evaluated existing models, but they were too slow for browser use.
  • Developed a custom “almost fully convolutional” network architecture, testing thousands of experimental variants via grid search for optimal hyperparameters.
  • Tackled ambiguity in gray-scale pixel classification by running A/B tests with human evaluators and built an objective accuracy metric to mirror human perception.
  • Created a proprietary dataset combining: internally generated media, public datasets, user-submitted videos, and publicly shared whiteboard recordings.
  • Resulted in the neural model lovingly called by us: “STB‑745”, optimized to recognize content and human presence reliably, no dedicated hardware required.

Work we’ve delivered

STB‑745 Neural Network Architecture
Custom semantic segmentation model tailored for browser inference at high accuracy
Neural network optimization
Light model
Light model
Custom Data Collection Pipeline
Built multi-source dataset including internal, public, prototype user videos
Dataset creation
Dataset creation
Training data
Training data
Training data
Training data
No items found.
No items found.

Tech Stack

Frontend
JavaScript
Custom STB‑745 convolutional neural network (trained offline)
PyTorch
Tensorflow
Backend
JavaScript
Custom STB‑745 convolutional neural network (trained offline)
PyTorch
Tensorflow
Platform
JavaScript
Custom STB‑745 convolutional neural network (trained offline)
PyTorch
Tensorflow
DevOps
JavaScript
Custom STB‑745 convolutional neural network (trained offline)
PyTorch
Tensorflow
Design
JavaScript
Custom STB‑745 convolutional neural network (trained offline)
PyTorch
Tensorflow
Other
JavaScript
Custom STB‑745 convolutional neural network (trained offline)
PyTorch
Tensorflow
JavaScript
Custom STB‑745 convolutional neural network (trained offline)
PyTorch
Tensorflow

Results

  • Built STB‑745 neural network from scratch in under 12 months.
  • Enables real-time segmentation in a web browser, compatible with most laptops and no need for native installs.
  • Achieves 96.3 % accuracy in content identification.
  • Identifies whiteboard content in under 1 second across devices, offering seamless UX.

Conclusion

In under a year, we created STB‑745: a high-accuracy, lightweight neural segmentation model tailored for real-time browser use. By combining custom architecture, extensive tuning, and unique data sources, we delivered a solution that makes ShareTheBoard scalable, accessible, and powerful. The outcome: fast, reliable content capture without hardware limitations, helping ShareTheBoard reach broader audiences and deliver on its mission.

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.

Dawid Skurzok

Dawid Skurzok

Machine Learning Engineer
Jerzy Nowiński

Jerzy Nowiński

Head of Engineering
Marcin (Martin) Demkowicz

Marcin (Martin) Demkowicz

Strategic Advisor

Case study

You’re the center of our work

Your satisfaction is our main metric. Join hundreds of clients from across the globe
and experience quality as a non-negotiable

TrustRadius: Navigating Digital Transformation
TrustRadius: Navigating Digital Transformation
WordPress
SEO
UX & UI Design
CRO
Fatty15 Shopify Transformation
Fatty15 Shopify Transformation
UX & UI Design
QA
WordPress
Shopify
Shopify
Running Neural Networks Inside the Web Browser? Only if You Know How!
Running Neural Networks Inside the Web Browser? Only if You Know How!
EdTech
Machine Learning
Computer Vision

Available for projects

Your next idea starts here!

Let’s partner up and build, grow,
and maintain your product together.

Thank you for reaching out! We'll get back to you shortly
Oops! Something went wrong while submitting the form.
Contact Ala
Alicja Trzewik-Wantuła
Director of Client Relations
‍