Developer Builds 6.4M-Parameter Transformer From Scratch to Power Recipe App

A developer built a custom decoder-only transformer called RasavedaGPT entirely from scratch using PyTorch, without relying on any external APIs or pretrained weights. The 6.4-million-parameter model powers Rasaveda, a recipe intelligence app that matches ingredients to recipes and explains cooking steps conversationally. Trained in two stages on a single Google Colab T4 GPU, the model first learned general English from WikiText-2 before being fine-tuned on recipe-specific data. At its small size, the model runs inference on CPU within a FastAPI backend, eliminating API costs, rate limits, and third-party dependencies. The project was driven by both practical self-sufficiency goals and a personal desire to deeply understand transformer architecture from the ground up.