I Made Tyre Size Detection App Using Gemma4:e4b
This is a submission for the Gemma 4 Challenge: Build with Gemma 4 What I...
*This is a submission for the [Gemma 4 Challenge: Build with Gemma 4](https://dev.to/challenges/google-gemma-2026-05-06)*
What I Built
<!-- Provide an overview of your project and what problem it solves or experience it creates. --> **NewTyre-AI** is a secure, localized full-stack web application designed to automate a deceptively complex industrial task: passenger vehicle tyre sidewall size extraction.
The Problem
Traditional computer vision pipelines frequently default to cloud-dependent architectures to process visual data. While convenient, this approach introduces persistent operational liabilities for businesses: recurring cloud API bills, data transit latencies, and corporate data privacy exposure. Furthermore, reading alphanumeric characters from a tyre sidewall is difficult for traditional linear Optical Character Recognition (OCR) tools because text printed on rubber is non-linear, low-contrast, heavily textured, and curved.
The Solution
NewTyre-AI solves this by shifting the entire visual processing workload onto a local, physical on-premise company server. The system ingests sidewall photos, filters them through a localized multimodal edge model, and returns an un-hallucinated, deterministic 9-character tyre size code (e.g., 205/60R16) to the technician with zero ongoing cloud compute or external API costs.
Demo
<!-- Embed a video walkthrough or share a link to your deployed project. --> Video Demo(30 sec): [Google Drive](https://drive.google.com/file/d/1sp0odXgkxL9jELnCiXovkT5TY1zEmCjg/view?usp=sharing)
Code
<!-- Embed or share a link to your repository. --> [GitHub Repo](https://github.com/inusha-thathsara/NewTyre-AI)
How I Used Gemma 4
<!-- Explain how Gemma 4 powers your project. Tell us which model you chose (E2B, E4B, or 31B Dense) and why it was the right fit for your use case. --> For this project, I deliberately selected the Gemma4:e4b (Effective 4B) parameter model rather than scaling up to the massive dense weights or downgrading to the highly lightweight e2b version. 2B model lack the visual cross-attention layer density required to accurately map characters in noisy geometric layouts, while 32B model is a overkill. The e4b model retains the precise structural transformer resolution required to read curved, dirty text on dark rubber cylinders without throwing a wave of false positives.
<!-- Don't forget to add a cover image if you want! -->
<!-- Team Submissions: Please pick one member to publish the submission and credit teammates by listing their DEV usernames directly in the body of the post. -->
<!-- Thanks for participating! -->