Product Engineer · Passion Project· June 2026
AgriScan AI is an offline-first mobile diagnostic app for Filipino corn farmers. It identifies three common corn diseases: Leaf Spot, Common Rust, and Northern Leaf Blight from a single photo in under three seconds, with no internet connection required. For farmers in Mindanao and the Visayas, where the nearest agronomist may be hours away, AgriScan provides an accurate, instant second opinion right in the field.
Smallholder corn farmers in the Philippines face preventable crop losses every season from diseases that are treatable if caught early. The problem is the gap between when a disease appears and when a farmer can get a reliable diagnosis.
The nearest agricultural extension officer is often more than two hours away. Laboratory sample results can take 5 to 10 days. Mobile data in farming areas is unreliable, and most farmers rely on visual guesswork that varies widely in accuracy.
By the time a farmer receives a correct diagnosis, the disease has already spread and crop yield has been lost.
Product Engineer
June 2026
AgriScan AI puts an on-device diagnostic model directly on the farmer's phone. Take a photo of the affected leaf and get a result in under three seconds. No internet, no account, and no waiting.
The app runs a MobileNet V2 TFLite model trained on the PlantVillage dataset entirely offline. Every scan is saved locally in a persistent history log, allowing farmers and agricultural extension officers to track disease occurrence over time.
Point your camera at a corn leaf and get an instant diagnosis, including Leaf Spot, Common Rust, Northern Leaf Blight, or Healthy, along with a confidence score.
The model was trained in Google Teachable Machine using 3,852 images across all four classes, then exported as a TFLite model that runs entirely on-device. No network call is ever made.
A 2×2 grid of preloaded sample images allows recruiters and technical evaluators to run the full TFLite diagnostic pipeline directly in a web browser through Appetize.io, with no physical device, installation, or setup required. Each tile triggers real on-device inference using a bundled image asset.
Every diagnostic result is savedautomatically to a local Hive database — no cloud, no sync, no accountrequired. The history log displays results in reverse-chronological order withlocalised timestamps, and persists across app restarts and OS kills.
Every UI decision was made for directsunlight, gloved hands, and interrupted attention. The oversized 116dp scan reticle, high-contrast #163300 / #9FE870 palette, and IndexedStack navigation all exist because this app is used in a field, not at a desk.
All app state — model loading status, scanresults, and history — is managed through Riverpod providers. The scanProviderhandles the full diagnostic lifecycle from image capture to result display,making the ML pipeline predictable, testable, and extensible.
AgriScan targets 2–3GB RAM Android devices— the real-world floor for smartphones in rural Philippine agriculture. Imagecompression is enforced at 85% quality before ML tensor allocation to preventout-of-memory crashes. The TFLite model was exported from Google TeachableMachine and selected specifically for its lightweight footprint on constrainedhardware.
AgriScan AI is deployed and available to try.
View Live DemoI designed and built AgriScan AI end to end, including model training in Google Teachable Machine, TFLite integration, Riverpod state management, UI design, and the Portfolio Sandbox Mode. The challenge was creating an agricultural tool that works entirely offline, with no internet, backend, or cloud services, while remaining fast and intuitive.
The Portfolio Sandbox Mode was designed so recruiters can experience the app in a browser without a physical device or camera access, while still running real on-device machine learning inference.
