MangIA: Smart Automated Classification Mango

What it does

MangIA classifies export-quality mangoes with 96.33% accuracy, eliminating human errors and speeding up the process 30x. It works offline, ideal for rural areas, and professionalizes an artisanal task.

Your inspiration

I observed Colombian farmers manually sorting mangoes for hours, facing fatigue and errors that affected exports. Colombia doubled its mango exports (2019-2024), but the methods remain artisanal. I wanted to democratize artificial intelligence for small producers, creating a system that captures decades of farming experience and consistently reproduces it without the need for the internet, accessible to anyone and with all the autonomy and precision of the best high-performance sorters, targeting the low- and medium-scale producer sector.

How it works

MangIA combines specialized hardware with edge artificial intelligence. The system includes a MaixCam smart camera with dedicated neural processor, operating in a controlled environment with radial LED lighting and white curved background. 1- Capture: Camera takes a photo of the mango under standardized conditions 2- Analysis: Custom CNN algorithm analyzes color, texture, and shape 3- Classification: In 70.3 milliseconds, identifies among 6 categories: unripe, optimal, ripe, diseased, other fruit, or no fruit 4- Result: Displays classification on touchscreen with confidence percentage The magic lies in the SimpleCNN algorithm I designed specifically for this limited hardware. With only 16 layers and 459KB size, it outperforms complex commercial architectures. It operates completely on-device, without external servers, processing all artificial intelligence locally with just 1.5W consumption.

Design process

I began researching international standards and visiting Colombian farms to understand the real manual process. I evaluated multiple hardware platforms, selecting MaixCam for its price-performance balance. I built a controlled capture environment and developed an application to collect 12,800 balanced images. I designed and trained my SimpleCNN algorithm from scratch, comparing it against 6 optimized architectures (MobileNet, EfficientNet, ResNet) and traditional methods. SimpleCNN emerged as winner through multi-criteria evaluation. I converted the model to MaixCam format, where quantization maintained 99.95% fidelity. Finally, I validated with 300 real samples against human experts, where MangIA consistently outperformed manual classification.

How it is different

1- Total Autonomy: First system operating 100% offline. Competitors require internet or expensive hardware. 2- Extreme Efficiency: 459KB and 70.3ms response, 177 times smaller than existing solutions while maintaining superior accuracy. 3- Specialized Algorithm: SimpleCNN designed specifically for mangoes, not generic adaptation. Outperforms MobileNetV3 (46.33%) and commercial MaixHub (84%). 4- Accessible Hardware: $66 vs $249 alternatives, democratizing advanced AI for small rural producers.

Future plans

I will expand MangIA by integrating industrial conveyor belts and developing weather-resistant versions for direct field use. I'll adapt the system for other tropical fruits and create a complete traceability mobile app. My vision is to license the technology globally, reducing post-harvest losses from current 30% to 5%.

Awards

1- Accepted: IFAC AGRICONTROL 2025 International Conference (California, USA) 2- Under review: Article for IEEE Transactions on AgriFood Electronics 3- Published: Colombian School of Engineering Journal 4- Under review: Paper to be a speaker at IEEE Café 2025 (Uruguay)