Replication Artifacts: Edge-Optimized 20M-Custom Vision-Language Model For OptiPros

This repository contains the deployment artifacts, quantized model graphs, and empirical field-testing logs for the custom model designed for OptiPros.

Primary Contribution

Our core work introduces a custom hardware-aware Vision Transformer (ViT) Encoder + Transformer Decoder pipeline trained on the standardized Microsoft COCO 2014 dataset. When optimized via 8-bit quantization, this architecture outperforms the state-of-the-art SmolVLM-256M baseline in both task accuracy and inference latency on edge hardware platforms.

Live Deployment Demonstration

A complete hardware-in-the-loop video demonstration highlighting both the physical hardware infrastructure and synchronized real-time software views while the device is actively worn by a user is publicly accessible:

🔗 Watch the Live OptiPros Device Demo Video Here

Empirical Field-Testing Validation

The terminal screenshots below document the live execution traces of our INT8 quantized model running on resource-constrained target edge hardware.

1. Static Environmental Scene Captioning

This trace validates the target hardware engine processing stationary structural elements, outputting a highly accurate contextual string mapping without frame lag.

2. Stationary Vehicle Asset Captioning

This live execution snapshot captures a stationary vehicle in a residential layout, verifying the ultra-low inference latency and high caption fidelity achieved by the custom architecture.

---

Repository Contents

• /models/: Contains the frozen, production-ready INT8 quantized computational graphs (encoder_int8.onnx and decoder_int8.onnx). Due to graph serialization and quantization adjustments, these files represent static inference models optimized for resource-constrained runtimes.
• /assets/: Contains field-testing terminal screenshots. These images capture the real-time console log streams on target hardware, verifying the operational performance detailed in the manuscript.

Terms of Use & Licensing

This repository is published strictly for academic peer-review and reproducibility verification. All model artifacts and logs are protected under the Creative Commons Attribution-NonCommercial (CC BY-NC 4.0) license. Commercial deployment or enterprise replication of these frozen weights is strictly prohibited.