Millnew AI – On device time-series prediction AI using PyTorch & Captum

RobinHood is an open-source, hybrid local framework designed to analyze multi-asset time-series data. It uses a deep learning architecture that combines TCN, LSTM, and Attention with a mathematical feature attribution layer based on Captum Integrated Gradients. The framework states that this combination is intended to eliminate hallucinations in automated financial reports.

The design separates local computation from text synthesis. Raw time-series data, deep learning training loops, and Captum mathematical feature attributions run entirely on the user's hardware. Only final, anonymized deterministic numerical weights are sent via API to open-source LLMs on Hugging Face to generate a human-readable text report, and the raw data never leaves the machine.

A ready-to-use gld_price_data.csv dataset is included for testing. For custom data, the system exclusively parses structured .csv files, evaluates standalone data matrices, and requires purely numerical time-series except for a single chronologically sequential date column. High data density is described as translating to robust mathematical attribution maps.