Inspiration
Informal retail feeds billions of people, yet most point-of-sale systems are designed for environments with stable internet, fixed product catalogs, and formal workflows. Street vendors already own smartphones, but the tools they need to turn those devices into practical business infrastructure do not exist. This project was inspired by the idea that the camera vendors already use every day could become the POS itself, if the underlying models are accurate, lightweight, and robust enough to run on-device.
What it does
EdgeVision POS turns a smartphone camera into an offline point-of-sale and inventory interface. Vendors photograph fruits and vegetables as they receive stock and again at the point of sale, and a compact vision model recognizes the produce locally without barcodes, typing, predefined SKUs, or internet connectivity. This project demonstrates how dendritic optimization can significantly improve recognition accuracy while keeping the model small enough for edge deployment.
How we built it
We fine-tuned a MobileNetV3-Small backbone on a fruit and vegetable recognition dataset and used this as a baseline. We then applied dendritic optimization using PerforatedAI’s library, integrating dendrites via the required APIs (initialize_pai() and add_validation_score()) without changing the underlying architecture. Training was run as a controlled comparison between the baseline and dendritic versions, using identical data splits, transforms, and optimization settings. The dendritic model adaptively restructured during training based on validation performance and produced the required PerforatedAI results graph.
Challenges we ran into
The primary challenge was integrating dendritic optimization in a constrained execution environment. Enabling deeper dendritic backpropagation led to environment-level execution issues, which required narrowing the implementation to supported integration paths. Ensuring reproducibility, non-interactive execution, and compatibility with Kaggle while still generating the mandatory PerforatedAI output artifacts also required careful configuration and iteration.
Accomplishments that we’re proud of
We achieved a substantial accuracy improvement on a compact, edge-friendly model without increasing the backbone size, eliminating approximately 47% of the baseline model’s remaining validation error. The project is fully reproducible, compliant with hackathon requirements, and grounded in a real-world deployment scenario rather than a synthetic benchmark. Most importantly, it demonstrates that dendritic optimization can meaningfully improve practical models intended for use outside controlled lab environments.
What we learned
Dendritic optimization is especially well-suited to compact architectures where traditional scaling is not an option. Validation-driven restructuring can extract additional performance without bloating parameter count, which is critical for edge deployment. We also learned the importance of tight experimental control and reproducibility when evaluating architectural changes, particularly in hackathon and review-driven settings.
What’s next for EdgeVision POS
Next steps include extending the system to incremental, on-device learning so models can adapt to each vendor’s unique inventory, integrating multimodal inputs such as audio for transaction capture, and building lightweight workflows for real-time inventory reconciliation. Longer term, this approach can expand beyond produce to other informal retail categories, further validating the camera-as-POS paradigm.
Log in or sign up for Devpost to join the conversation.