AgriCuke: AI-Powered Cucumber Farming Assistant

Project Overview

AgriCuke is an AI-powered farming assistant built for cucumber growers who need fast, reliable answers in the field. Farmers photograph a leaf or a section of the crop and receive an on-device disease classification, a treatment plan, and a soil and climate reading log that syncs to the cloud for season-long analytics.

Cucumber cultivation is unusually sensitive to early leaf disease, humidity swings, and soil moisture drift. A missed diagnosis in the first 48 hours can wipe out weeks of growth. AgriCuke closes that reaction window by combining computer vision, real-time telemetry, and simple mobile UX so growers act on evidence rather than guesswork.

Engineering Challenges Solved

• On-Device Disease Classification: Cucumber leaf conditions such as powdery mildew, downy mildew, and bacterial wilt look similar to the untrained eye. A TensorFlow model was trained on labeled leaf imagery and quantized for on-device inference so diagnosis works offline in rural fields with unstable connectivity.
• Real-Time Field Telemetry: Soil moisture, ambient temperature, and growth stage readings needed to reach the cloud without draining phone battery. The Flutter client batches captures and syncs opportunistically to Firebase Realtime Database whenever a connection returns.
• Farmer-First Mobile UX: Growers were the primary users, not agronomists. The interface reduces every action to a single tap, uses large tap targets suitable for gloved hands, and speaks in plain language rather than lab terminology.

Technical Architecture and Implementation

Mobile Client Application (Flutter)

The app was designed in Figma and built with Flutter for uniform performance across iOS and Android. Core flows include a leaf scan capture with viewfinder guidance, a daily field reading log for soil moisture and temperature, a growth stage tracker, and a personalized recommendation feed.

AI Inference Layer (TensorFlow, Python)

A convolutional neural network trained in Python on a labeled cucumber leaf dataset classifies photos into healthy, powdery mildew, downy mildew, bacterial wilt, and nutrient deficiency categories. The model was converted to TensorFlow Lite for on-device execution, keeping inference under two seconds on mid-range Android phones.

Cloud Sync and Recommendations (Firebase)

Firebase Authentication scopes each grower's data to their farm profile. Firebase Realtime Database stores field readings, diagnosis history, and treatment progress, while Firebase Cloud Messaging pushes seasonal alerts such as high humidity warnings and irrigation reminders.

Measurable Outcomes

• Under 2s On-Device Diagnosis: Leaf classification runs locally on mid-range Android hardware with no server round trip required.
• 90 Percent Model Accuracy in Field Testing: Verified across a pilot cohort of cucumber growers on real, non-training-set leaf samples.
• Offline First: All captures and readings queue locally and sync automatically when the device regains connectivity, so field work is never blocked by signal loss.
• Unified Codebase: A single Flutter codebase serves both iOS and Android, cutting delivery time by roughly half compared with parallel native builds.