Leaf Inference Engine

Farmers needed a fast, low-latency way to diagnose plant disease from a photo, without specialized agricultural equipment or reliable network access in rural areas.

A CNN trained on roughly 87,000 leaf images across 38 disease classes reaches about 95% accuracy on held-out images. The Next.js frontend never talks to the model directly; it goes through a proxy that keeps the FastAPI endpoint and its keys off the client.

Training happens offline and produces a standard .h5 model file. In production, the Next.js server proxy holds the API keys and forwards requests to a FastAPI service running on Railway, which loads that Keras model into memory and serves inference from it.

A heuristic check rejects non-leaf images before inference runs, so a photo of a hand or a wall doesn't come back as a confident wrong diagnosis. A Gemini-powered companion then suggests organic treatment options for whatever gets identified.

1. 01

   Proxy Security Layer

   The FastAPI endpoints and API keys never reach the browser. The Next.js API layer forwards multipart form data to Railway, so a client can't hit the inference endpoint directly or see what key it's using.

2. 02

   Single Source Data Pipeline

   config.py and data_pipeline.py are the single source for preprocessing constraints, used both when the dataset gets compiled and when a live image gets validated. The same resize and normalization logic runs in training and production, so there's no drift between the two.