DabbaETA: a 'will this tiffin be late?' predictor that never sends a customer's address to the cloud

DabbaETA is a local-first assistant designed to help small tiffin, bakery, and courier businesses accurately predict delivery times and identify orders at risk of delays. The tool allows owners to upload past delivery data in CSV format, which is then stored locally in a SQLite file. Using this historical data, DabbaETA predicts the ETA and late-delivery risk for each pending order, displaying the results as HIGH, MODERATE, or LOW risk levels.

The core of DabbaETA lies in its prediction engine, called TabPFN, which runs on the local machine. TabPFN uses machine learning models to analyze factors such as distance, traffic, weather, vehicle type, preparation time, courier experience, evening rush, and monsoon effects to generate accurate ETA estimates. Gemma, powered by the Ollama language model, assists in drafting polite delay messages in Hinglish, Hindi, or English, ready to be copied into WhatsApp conversations with customers.

DabbaETA operates entirely offline, ensuring that customer data remains private and never leaves the owner's machine. Instead, only a table of numbers containing relevant metrics is sent to TabPFN for processing. The tool is highly customizable, allowing owners to easily swap in better small Hindi-capable models as they become available by simply updating an environment variable.

This flexibility ensures that customers do not face lock-in with a specific vendor, as they can adapt the model to their needs without affecting daily operations.